Predicting accurate solid-state thermodynamics is easier than you think
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Condensed Matter. The day in condensed matter: materials, phases and the measurements under them.
Kai: I'm Kai.
Mira: And I'm Mira, here with Lev.
Kai: We're looking at the paper: "Predicting accurate solid-state thermodynamics is easier than you think".
Mira: Here's the gist of the paper. Predicting challenging phase transitions with Bayesian active learning enables an on-the-fly framework combining the stochastic self-consistent harmonic approximation and Bayesian ML potentials to accelerate first-principles accurate finite temperature calculations, which can lower the SSCHA first-principles workload by 98.4 – 99.8%.
Kai: This method is shown to accurately capture challenging phase transitions in materials like Li2O and CsPbI3 with remarkable efficiency.
Mira: Materials Context Materials underpin modern technologies, from energy harvesting, storage, and conversion to information and communication technologies.
Kai: The functionality of these materials is often governed by the interplay between competing phases, as thermodynamic behavior shapes microscopic properties and ultimately determines technological performance.
Paper summary: Kai: So this paper is called "Predicting accurate solid-state thermodynamics is easier than you think", and it’s about using a specific combination of methods to speed up those calculations.
Mira: It's looking at how we can handle challenging phase transitions in materials, which usually requires a lot of expensive first-principles work. The paper claims this new setup makes that process much faster, specifically by lowering the workload by ninety-eight point four to ninety-nine point eight percent.
Lev: That’s a huge reduction in computational time when you're thinking about how many DFT calculations you have to run for finite temperature physics <ref:2604.25756#pg2>. I wonder if that massive speedup translates well when we take it to real hardware, though.
Kai: Well, the core idea they present in "Predicting accurate solid-state thermodynamics is easier than you think" is combining the stochastic self-consistent harmonic approximation with Bayesian machine learning potentials. They use this framework to accelerate calculations for first-principles accurate finite temperature results <ref:2604.25756#pg2>.
Mira: What really matters here, as they point out, is that this approach can efficiently apply to a wide range of temperatures and accurately capture challenging phase transitions <ref:2604.25756#pg2>. They show this on materials like Lithium Oxide, Li2O, and Cesium Lead Iodide, CsPbI3.
Lev: So they’re applying it to things that have complex behavior, like the competition between black and yellow phases in CsPbI3 <ref:2604.25756#pg2>. That complexity is what makes the standard harmonic or quasi-harmonic approximations fall short, which is a key motivation here six <ref:2604.25756#pg1>.
Kai: Exactly. The paper shows that this on-the-fly active learning scheme works by generating an ensemble of structures from a nuclear Gaussian density matrix and then selecting them based on the uncertainty from a sparse Gaussian process force field <ref:2604.25756#pg3>.
Mira: Then they label those selected structures using DFT to get energies, forces, and stress, which they then use to update the machine learning potential <ref:2604.25756#pg3>. That iterative process is what makes it active learning.
Lev: From an engineering standpoint, I’m thinking about how that uncertainty threshold is set for selecting structures; they use a threshold of zero point zero five for structure selection and zero point zero zero five to add the sparse environments <ref:2604.25756#pg3>. That seems like a reasonable trade-off between accuracy and efficiency.
Kai: And the results they get are pretty direct, showing that this method accurately predicts the linear thermal expansion coefficient for Lithium Oxide, getting two point eight times ten to the negative five Kelvin per degree Celsius at one thousand Kelvin <ref:2604.25756#pg2>. That is in good agreement with what we see experimentally.
Mira: It’s also impressive how they handled CsPbI3, where they accurately captured the phase diagram, predicting a transition temperature of about five hundred sixty seven Kelvin for the yellow to black phase change <ref:2604.25756#pg2>. That was within thirty Kelvin of the experimental reference value which was five hundred ninety three Kelvin.
Paper summary: Lev: Those numbers are significant because they show that this method isn't just theoretical; it’s actually getting results that line up with what we expect from a material undergoing a phase transition <ref:2604.25756#pg3>. But I have to ask, how does the workload reduction hold up if we move away from these specific materials?
Kai: The paper points out that this scheme demonstrates a massive reduction in the first-principles computational workload required for thermodynamic calculations in crystalline materials, achieving ninety-eight point four to ninety-nine point eight percent reduction compared to reference SSCHA and DFT calculations <ref:2604.25756#pg2>.
Mira: That level of efficiency is what makes it so important because the underlying physics is contained in only a few configurations of the phase space, which means we don't need to sample everything <ref:2604.25756#pg2>.
Lev: So if I were trying to run this on real hardware for something like battery electrolytes, would that reduction translate directly into usable time savings on a standard cluster? The paper doesn't explicitly detail the hardware overhead, which is where things get tricky <ref:2604.25756#pg3>.
Kai: That’s a valid concern, Lev. The paper relies on the FLARE code for the Bayesian potential and uses quantum espresso v7 point two for DFT calculations <ref:2604.25756#pg3>. They have made it open source and available in the python-sscha code nine, which is good for accessibility <ref:2604.25756#pg2>.
Mira: And they are making sure the data is FAIR, meaning it’s available on the Materials Cloud Archive after publication <ref:2604.25756#pg3>. That commitment to reproducibility seems necessary when you’re dealing with these kinds of complex models.
Lev: It does sound like a solid approach for accelerating research right now, especially since they show this method can be used across a wide range of temperatures <ref:2604.25756#pg2>. But the paper does admit that the accuracy hinges on how good that machine learning potential is; if it’s not accurate enough, the dynamics might become unphysical <ref:2604.25756#pg3>.
Kai: That's a fair limitation they mention regarding the ML potential quality. It also has to be noted that they used a specific sampling method, an "optimal-guess" strategy for carrying out SSCHA calculations over a temperature range from an initial temperature T zero to a final temperature T max <ref:2604.25756#pg3>.
Mira: So the main thing to remember is that this paper shows you how combining these two tools—SSCHA and Bayesian ML potentials—can significantly speed up finite temperature calculations for solid-state thermodynamics, but you still have to trust the accuracy of your machine learning potential <ref:2604.25756#pg3>.
Lev: So it’s a powerful accelerator that lets you explore phase transitions faster, but it’s not a replacement for getting the underlying physical models right; you still need good input data or a very robust model <ref:2604.25756#pg3>.
Kai: That seems to be the main point of "Predicting accurate solid-state thermodynamics is easier than you think," showing that active learning can make exploring these materials much more efficient, even if we still have to validate the ML parts rigorously <ref:2604.25756#pg3>.
Conclusion: Kai: So we’ve been talking about how they used this active learning setup to speed up phase transition calculations.
Mira: Yeah, the paper is titled "Predicting accurate solid-state thermodynamics is easier than you think".
Kai: It’s about using a specific combination of methods—the stochastic self-consistent harmonic approximation and Bayesian machine learning potentials—to tackle those hard finite temperature problems.
Mira: They use this to get first-principles results that are much faster, cutting the workload by nearly a hundred percent for things like Lithium Oxide and Cesium Lead Iodide.
Lev: That ninety-eight point four to ninety-nine point eight percent reduction in first-principles workload is a big deal when you think about how much compute time you actually need on real hardware.
Kai: Exactly. The core idea is that the relevant physics in these phase transitions isn't spread out everywhere, it’s concentrated in just a few configurations of the phase space.
Mira: And they showed this works for predicting things like the thermal expansion coefficient for Lithium Oxide, getting results that match what we actually see experimentally at one thousand Kelvin.
Lev: That accuracy is impressive because it means we don't have to run those massive DFT calculations every single time we want to explore a temperature range.
Kai: It’s about making exploration efficient rather than just brute-forcing the whole phase space.
Mira: So what this paper really suggests is that by cleverly using active learning, we can get accurate thermodynamic data for materials much quicker than before.
Lev: It means you can test more material candidates or explore more temperature regimes in a much shorter time frame for the same computational cost.
Kai: It shifts the focus from just calculating a single point to intelligently selecting the most informative calculations based on what the machine learning model is uncertain about.
Mira: This method is designed to handle complex competition between different phases, like in CsPbI3, where you have black and yellow phases fighting for control.
Lev: The caveat we have is that the whole thing relies heavily on how well that initial machine learning potential is trained; if the model gets it wrong, then the active learning process just wastes time sampling bad structures.
Kai: Still, even with those caveats, they’ve shown a path to significantly accelerating our understanding of materials science thermodynamics.
Mira: It’s a step toward making high-accuracy finite temperature simulations more accessible for the next generation of material design.
Lorenzo Bastonero, Gabriel Joalland, § Chiara Cignarella, § Lorenzo Monacelli, Nicola Marzari
University of Bremen Excellence Chair, Bremen Center for Computational Materials Science, and MAPEX Center for Materials and Processes, University of Bremen · Theory and Simulation of Materials (THEOS), National Centre for Computational Design and Discovery of Novel Materials (MARVEL), École Polytechnique Fédérale de Lausanne · University of Rennes, École Nationale Supérieure de Chimie de Rennes, CNRS, Institut des Sciences Chimiques de Rennes
cond-mat.mtrl-sci, cond-mat.dis-nn, physics.comp-ph
Submitted: 2026-04-28
Updated: 2026-11-06
Comments: 10 main pages, 3 main figures, 6 supplementary pages, 9 supplementary figures
Code: https://github.com/SSCHAcode/python-sscha
License: http://creativecommons.org/licenses/by-nc-nd/4.0/
Importance score: 92/100
The gist: Predicting challenging phase transitions with Bayesian active learning enables an on-the-fly framework combining the stochastic self-consistent harmonic approximation and Bayesian ML potentials to…
Key concepts
- SSCHA
- Stochastic Self-Consistent Harmonic Approximation is a method used to calculate thermodynamic properties at finite temperatures. It helps describe the behavior of materials near phase transitions by considering vibrational modes, which are crucial for understanding how materials change structure when heated or cooled.
- Bayesian ML Potentials
- These are machine learning models that learn the potential energy surface of a material using data from first-principles calculations. They use Bayesian methods to quantify their uncertainty, allowing the system to select the most informative structures for further calculation, making the process more efficient.
- Active Learning Workflow
- This is an iterative strategy where a computational model (the ML potential) queries new data (DFT calculations) strategically. Instead of calculating everything, it focuses on structures where its prediction is uncertain, thereby learning the most important physics much faster and reducing the total required calculations.
- Phase Transition Analysis
- This involves studying how a material changes its structure when subjected to temperature changes. The method uses the Hessian of the positional free energy to monitor how vibrational modes change with temperature, allowing it to accurately predict transition temperatures between different phases.
Terminology
Summary
Predicting challenging phase transitions with Bayesian active learning enables an on-the-fly framework combining the stochastic self-consistent harmonic approximation and Bayesian ML potentials to accelerate first-principles accurate finite temperature calculations, which can lower the SSCHA first-principles workload by 98. p. 1 In
fact, an approximately 96% reduction in first-principles
workload has already been reported for PdCuH2 [26] at
fixed lattice geometry and temperature by pairing active
learning moment tensor potentials with SSCHA. In this work, we present an on-the-fly active-learning
scheme combining the SSCHA method [9] and Bayesian
ML potentials [19, 23] to greatly accelerate first-principle
accurate finite temperature calculations. This method
can be efficiently applied to a wide range of temperatures and accurately captures challenging phase transitions. With this approach, we showcase here two interesting energy materials with prototypical phase transitions with implications in solid-state batteries and photovoltaic research: Li2O and CsPbI3.Where the paper says it
Where the paper says it
Materials Context
Materials underpin modern technologies, from energy harvesting, storage, and conversion to information and communication technologies. The functionality of these materials is often governed by the interplay between competing phases, as thermodynamic behavior shapes microscopic properties and ultimately determines technological performance. p. 0 Predicting challenging phase transitions with Bayesian active learning
Lorenzo Bastonero ,
1, ∗ Gabriel Joalland ,
2, 3 Chiara
Cignarella ,
1, 2 Lorenzo Monacelli ,
2, 4 and Nicola Marzari 1, 2, 5, 6
1U Bremen Excellence Chair, Bremen Center for Computational Materials Science, and
MAPEX Center for Materials and Processes, Universität Bremen, 28359 Bremen, Germany
2Theory and Simulation of Materials (THEOS), and National Centre for Computational Design and Discovery
of Novel Materials (MARVEL), École Polytechnique Fédérale de Lausanne, 1015 Lausanne, Switzerland
3Univ Rennes, ENSCR, CNRS, Institut des Sciences Chimiques de Rennes, UMR 6226, Rennes, France
4Departement of Physics, Sapienza Università di Roma, Rome, Italy
5Theory of Condensed Matter, Cavendish Laboratory, University of Cambridge, Cambridge, CB3 0US, United Kingdom
6Laboratory for Materials Simulations, Paul Scherrer Institut (PSI), 5232 Villigen, Switzerland
Materials underpin modern technologies, from energy harvesting, storage, and conversion to information and communication technologies. Their functionality is often governed by the interplay between competing phases, as thermodynamic behavior shapes microscopic properties and ultimately
determines technological performance; for instance, the light absorption of inorganic metal-halide
perovskites in solar cells. This is the case, for example, of liquid-like
ionic mobility in superionic conductors for Li-ion batteries [1, 2], optoelectronic response of halide perovskites
for energy harvesting [3], and lattice thermal transport
of thermoelectrics for waste heat management [4, 5]. Moreover, many scientifically and technologically relevant compounds sit close to phase transitions, driven
by unstable vibrational modes [6], strong anharmonicity [4], and even non-negligible nuclear quantum fluctuations in light-element systems [7]. Moreover, many scientifically and technologically relevant compounds sit close to phase transitions, driven
by unstable vibrational modes [6], strong anharmonicity [4], and even non-negligible nuclear quantum fluctuations in light-element systems [7]. In these regimes, the
commonly-used harmonic and quasi-harmonic approximations are inadequate to describe the thermodynamic
driving forces [7–10].Where the paper says it
Where the paper says it
Where the paper says it
On-the-fly SSCHA Active Learning Workflow
The proposed on-the-fly active learning scheme combines the SSCHA method [9] and Bayesian ML potentials [19, 23] to greatly accelerate first-principles accurate finite temperature calculations. The workflow involves several key components:
-
Generation of an ensemble of structures from the nuclear Gaussian density matrix ρRi,Φi at iteration i. p. 2
Where the paper says it
Schematic overview of the presented on-the-fly active learning scheme. Top panels, left to right, the proposed active learning scheme involves: (a) the generation of an ensemble of structures from the nuclear Gaussian density matrix ρRi,Φi at iteration i; (b) the selection of structures based on ML uncertainty, using a sparse Gaussian process force field; (c) DFT labeling, i. p. 1
Where the paper says it
1(a) and (d)) by generating for each structure a vector of random numbers yµ and employing it as follows [28]: ra = Ra + X µ 1 √ Ma yµ s (2nµ + 1) ℏ 2ωµ e a µ . (3) As the labeling represents the most computationally demanding step of the SSCHA method, the main objective is to substitute the first-principles labeling, typically from density-functional theory (DFT) calculations, with a computationally inexpensive ML potential.
-
Selection of structures based on ML uncertainty, using a sparse Gaussian process force field. p. 2
Where the paper says it
Schematic overview of the presented on-the-fly active learning scheme. Top panels, left to right, the proposed active learning scheme involves: (a) the generation of an ensemble of structures from the nuclear Gaussian density matrix ρRi,Φi at iteration i; (b) the selection of structures based on ML uncertainty, using a sparse Gaussian process force field; (c) DFT labeling, i.
-
DFT labeling, i.e., perform ground-state calculation to assign total energy E, forces, and stress of the selected structures. p. 2
Where the paper says it
Top panels, left to right, the proposed active learning scheme involves: (a) the generation of an ensemble of structures from the nuclear Gaussian density matrix ρRi,Φi at iteration i; (b) the selection of structures based on ML uncertainty, using a sparse Gaussian process force field; (c) DFT labeling, i. , perform ground-state calculation to assign total energy E, forces, and stress of the selected structures, and use these to update the ML potential.
-
Update the SGP model with these labels. p. 1
Where the paper says it
To trade off between the number of unnecessary DFT calculations and the number of re-training calculations and the time-to-solution, we divide the ensemble into N batches, each containing Nc/N configurations, where Nc is the total number of structures in each ensemble. For each batch, we evaluate the uncertainty of the ML and select the uncertain structures, label them using DFT, and then update the SGP model with these labels (blue and orange panels in Fig.
Efficiency Gains in Workload Reduction
The scheme demonstrates a massive reduction of 2-3 orders of magnitude in the first-principles computational workload required for thermodynamic calculations in crystalline materials. For Li2O, only 44 DFT calculations were required to fit a reliable ML potential, which is comparable to the number needed for a quasi-harmonic approximation calculation. p. 3 Moreover, it
is worth noting that the 44 DFT calculations required for
the active-learning of Li2O are comparable in number to
those needed for a quasi-harmonic approximation calculation (see Supplementary Discussion), which, however,
does not capture full anharmonic effects. All these considerations indicate that SSCHA, coupled
with the active-learning scheme, enables a targeted exploration of the potential energy surface. The number of DFT labeling generally decreases with increasing SSCHA steps and temperatures. Notably, DFT calls are overall much fewer than the total number of configurations generated for each SSCHA
step, which are 200 and 400 for Li2O and α–CsPbI3, respectively. In
fact, an approximately 96% reduction in first-principles
workload has already been reported for PdCuH2 [26] at
fixed lattice geometry and temperature by pairing active
learning moment tensor potentials with SSCHA. In this work, we present an on-the-fly active-learning
scheme combining the SSCHA method [9] and Bayesian
ML potentials [19, 23] to greatly accelerate first-principle
accurate finite temperature calculations. All these considerations indicate that SSCHA, coupled
with the active-learning scheme, enables a targeted exploration of the potential energy surface. This allows not
only a highly efficient generation of the ML potential, but
also proves that the relevant physics is contained in only
a few configurations of the phase space.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Predictive Accuracy and Phase Transition Analysis
The framework successfully predicted (i) the linear thermal expansion coefficient of Li2O, yielding 2.8 × 10−5 K−1 at T0 = 1000 K, which is in excellent agreement with experimental values. p. 4 4, αT0
results in 2. 8 × 10−5 K−1 at T0 = 1000 K, which is in excellent agreement with the experimental values of about
3 × 10−5 K−1
[36, 37]. 8% while reproducing with high fidelity
the ab-initio reference thermodynamic results. Second,
we apply the method to predict (i) the linear thermal
expansion coefficient of Li2O, and (ii) the experimental
phase diagram of CsPbI3, where the spontaneous competition between the black (α) and yellow (δ) phases is
crucial in determining the light absorption properties of
the material. 3(b), and corresponds
to about 567 K, which is less than 30 K apart from the
experimental reference value Texp = 593 K [32]. Such a
small error indicates that SSCHA, combined with a good
model of the potential energy surface, provides a highfidelity description of the solid-solid phase transition.Where the paper says it
Where the paper says it
Where the paper says it
Methodological Details and Validation
The method employs an The eigenvalues of this Hessian
matrix tend towards zero as the system deviates from
the equilibrium position Req. This property enables the
study of second-order phase transitions by inspecting the
evolution of eigenvalues with temperature to determine
the critical transition temperature. Sparse Gaussian process (SGP)
We use the SGP force field implemented in the
FLARE code [19, 23]. For the particular application
shown in this study, we use the B(2) ACE [67] descriptors with nmax = 8 and 12 Chebyshev polynomials,
respectively for Li2O and CsPbI3, a maximum angular momentum lmax = 4, and a radial cutoff of 5.optimal-guess
strategy to carry out SSCHA calculations in a ramping temperature range, starting from an initial temperature T0, until a final temperature Tmax. The Hessian of the positional free energy F(R) is used to study second-order phase transitions by inspecting the evolution of eigenvalues with temperature. p. 6Where the paper says it
Where the paper says it
Data Availability and Open Source
The implementation is open source and available in the python-sscha [9] code, relying on the FLARE code [19, 23] for the Bayesian potential. The data used to produce the results of this work will be available on the Materials Cloud Archive after publication. p. 7 2, with a temperature step of ∆T = 50 K, and using the “optimal-guess”
strategy (see Optimal guess strategy section in Methods). DATA AVAILABILITY
The data used to produce the results of this work will
be available on the Materials Cloud Archive after publication. DATA AVAILABILITY
The data used to produce the results of this work will
be available on the Materials Cloud Archive after publication. CODE AVAILABILITY
All the code developed and presented here is
open source and made available on GitHub as
part of the official python-sscha distribution
(https://github. In this way, it could provide a unique and
self-contained framework for accelerated materials characterization. The implementation is open source and available in
the python-sscha [9] code, and relies on the FLARE
code [19, 23] for the Bayesian potential, and on the AiiDA
infrastructure [49–51] for automated and reproducible
calculations, following the FAIR principles [52].Where the paper says it
Where the paper says it
Where the paper says it
Computational Details Summary
DFT calculations are performed using Quantum-ESPRESSO distribution v7.2 [53–55]. The SSCHA calculations are performed using a branch of python-sscha code v1. p. 6 All calculations are submitted and automatically managed by the aiida-quantumespresso [50, 51]
and aiida-quantumespresso-ph plugins of the AiiDA
framework [50, 51]. Self-consistent stochastic approximation (SSCHA)
The SSCHA calculations are performed using a branch
of python-sscha code v1. Self-consistent stochastic approximation (SSCHA)
The SSCHA calculations are performed using a branch
of python-sscha code v1. 1 with SHA git commit
ed837e0a1d4ad39a7500554117352ebbd932b0f5. Density-functional perturbation theory (DFPT)
calculations, used to obtain the starting dynamical matrices, dielectric and Born effective charge tensors, were
performed using the PHonon code [65] of Quantum
ESPRESSO using a q-point mesh of 2×2×2, 2×2×2,
and 1 × 2 × 1 respectively for Li2O, α-CsPbI3, and δCsPbI3. For the SSCHA calculations reported in Fig. The SSCHA calculations reported in Fig. 2 were run
in 2×2×2, 2×2×2, and 1×2×1 supercells, for Li2O, αCsPbI3, and δ-CsPbI3, respectively, commensurate to the
q-point meshes of the DFPT calculations for harmonic
dynamical matrices. For the particular application
shown in this study, we use the B(2) ACE [67] descriptors with nmax = 8 and 12 Chebyshev polynomials,
respectively for Li2O and CsPbI3, a maximum angular momentum lmax = 4, and a radial cutoff of 5. We use a normalized dot-product kernel with an exponent ξ = 2. 005 for the threshold to add the
sparse environments (corresponding to 10% of the uncertainty threshold). We tried different values of the uncertainty thresholds for structure selection, and found that
0. We tried different values of the uncertainty thresholds for structure selection, and found that
0. 05 provides a good trade-off between accuracy and efficiency (see Supplementary Figure 9). Each ensemble was
subdivided into batches containing 8 configurations each. The active-learning simulations are carried out over the
entire temperature range shown in Fig. Each ensemble was
subdivided into batches containing 8 configurations each. The active-learning simulations are carried out over the
entire temperature range shown in Fig. Each ensemble was
subdivided into batches containing 8 configurations each. The active-learning simulations are carried out over the
entire temperature range shown in Fig. Each ensemble was
subdivided into batches containing 8 configurations each. The active-learning simulations are carried out over the
entire temperature range shown in Fig. Each ensemble was
subdivided into batches containing 8 configurations each. The active-learning simulations are carried out over the
entire temperature range shown in Fig. Each ensemble was
subdivided into batches containing 8 configurations each. The active-learning simulations are carried out over the
entire temperature range shown in Fig. Each ensemble was
subdivided into batches containing 8 configurations each. The active-learning simulations are carried out over the
entire temperature range shown in Fig. Each ensemble was
subdivided into batches containing 8 configurations each. The active-learning simulations are carried out over the
entire temperature range shown in Fig. Each ensemble was
subdivided into batches containing 8 configurations each. The active-learning simulations are carried out over the
entire temperature range shown in Fig. Each ensemble was
subdivided into batches containing 8 configurations each. The active-learning simulations are carried out over the
entire temperature range shown in Fig. Each ensemble was
subdivided into batches containing 8 configurations each. The active-learning simulations are carried out over the
entire temperature range shown in Fig. Each ensemble was
subdivided into batches containing 8 configurations each. The active-learning simulations are carried out over the
entire temperature range shown in Fig. Each ensemble was
subdivided into batches containing 8 configurations each. The active-learning simulations are carried out over the
entire temperature range shown in Fig. Each ensemble was
subdivided into batches containing 8 configurations each. The active-learning simulations are carried out over the
entire temperature range shown in Fig. Each ensemble was
subdivided into batches containing 8 configurations each. The active-learning simulations are carried out over the
entire temperature range shown in Fig. Each ensemble was
subdivided into batches containing 8 configurations each. The active-learning simulations are carried out over the
entire temperature range shown in Fig. Each ensemble was
subdivided into batches containing 8 configurations each. The active-learning simulations are carried out over the
entire temperature range shown in Fig. Each ensemble was
subdivided into batches containing 8 configurations each. The active-learning simulations are carried out over the
entire temperature range shown in Fig. Each ensemble was
subdivided into batches containing 8 configurations each. The active-learning simulations are carried out over the
entire temperature range shown in Fig. Each ensemble was
subdivided into batches containing 8 configurations each. The active-learning simulations are carried out over the
entire temperature range shown in Fig. Each ensemble was
subdivided into batches containing 8 configurations each. The active-learning simulations are carried out over the
entire temperature range shown in Fig. Each ensemble was
subdivided into batches containing 8 configurations each. The active-learning simulations are carried out over the
entire temperature range shown in Fig. Each ensemble was
subdivided into batches containing 8 configurations each. The active-learning simulations are carried out over the
entire temperature range shown in Fig. Each ensemble was
subdivided into batches containing 8 configurations each. The active-learning simulations are carried out over the
entire temperature range shown in Fig. Each ensemble was
subdivided into batches containing 8 configurations each. The active-learning simulations are carried out over the
entire temperature range shown in Fig. Each ensemble was
subdivided into batches containing 8 configurations each. The active-learning simulations are carried out over the
entire temperature range shown in Fig. Each ensemble was
subdivided into batches containing 8 configurations each. The active-learning simulations are carried out over the
entire temperature range shown in Fig. Each ensemble was
subdivided into batches containing 8 configurations each. The active-learning simulations are carried out over the
entire temperature range shown in Fig. Each ensemble was
subdivided into batches containing 8 configurations each. The active-learning simulations are carried out over the
entire temperature range shown in Fig. Each ensemble was
subdivided into batches containing 8 configurations each. The active-learning simulations are carried out over the
entire temperature range shown in Fig. Each ensemble was
subdivided into batches containing 8 configurations each. The active-learning simulations are carried out over the
entire temperature range shown in Fig. Each ensemble was
subdivided into batches containing 8 configurations each. The active-learning simulations are carried out over the
entire temperature range shown in Fig. Each ensemble was
subdivided into batches containing 8 configurations each. The active-learning simulations are carried out over the
entire temperature range shown in Fig. Each ensemble was
subdivided into batches containing 8 configurations each. The active-learning simulations are carried out over the
entire temperature range shown in Fig. Each ensemble was
subdivided into batches containing 8 configurations each. The active-learning simulations are carried out over the
entire temperature range shown in Fig. Each ensemble was
subdivided into batches containing 8 configurations each. The active-learning simulations are carried out over the
entire temperature range shown in Fig. Efficient and accurate thermodynamic properties
First, we use the method presented to calculate
the thermodynamic properties of Li2O and α–CsPbI3,
demonstrating the efficiency of the proposed approach. In Figures 2(a) for Li2O and (d) for α-CsPbI3, the number of DFT calculations is shown as a function of the
SSCHA minimization step and temperature. On-the-fly active learning
The active learning workflow presented in this work
is made available in the official python-sscha code as a
part of a new class called AiiDAEnsemble [16], which also
handles the automated submission of DFT jobs. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. On-the-fly active learning
The active learning workflow presented in this work
is made available in the official python-sscha code as a
part of a new class called AiiDAEnsemble [16], which also
handles the automated submission of DFT jobs. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. On-the-fly active learning
The active learning workflow presented in this work
is made available in the official python-sscha code as a
part of a new class called AiiDAEnsemble [16], which also
handles the automated submission of DFT jobs. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. On-the-fly active learning
The active learning workflow presented in this work
is made available in the official python-sscha code as a
part of a new class called AiiDAEnsemble [16], which also
handles the automated submission of DFT jobs. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. On-the-fly active learning
The active learning workflow presented in this work
is made available in the official python-sscha code as a
part of a new class called AiiDAEnsemble [16], which also
handles the automated submission of DFT jobs. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. On-the-fly active learning
The active learning workflow presented in this work
is made available in the official python-sscha code as a
part of a new class called AiiDAEnsemble [16], which also
handles the automated submission of DFT jobs. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. On-the-fly active learning
The active learning workflow presented in this work
is made available in the official python-sscha code as a
part of a new class called AiiDAEnsemble [16], which also
handles the automated submission of DFT jobs. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. On-the-fly active learning
The active learning workflow presented in this work
is made available in the official python-sscha code as a
part of a new class called AiiDAEnsemble [16], which also
handles the automated submission of DFT jobs. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. On-the-fly active learning
The active learning workflow presented in this work
is made available in the official python-sscha code as a
part of a new class called AiiDAEnsemble [16], which also
handles the automated submission of DFT jobs. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. On-the-fly active learning
The active learning workflow presented in this work
is made available in the official python-sscha code as a
part of a new class called AiiDAEnsemble [16], which also
handles the automated submission of DFT jobs. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. On-the-fly active learning
The active learning workflow presented in this work
is made available in the official python-sscha code as a
part of a new class called AiiDAEnsemble [16], which also
handles the automated submission of DFT jobs. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. On-the-fly active learning
The active learning workflow presented in this work
is made available in the official python-sscha code as a
part of a new class called AiiDAEnsemble [16], which also
handles the automated submission of DFT jobs. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. On-the-fly active learning
The active learning workflow presented in this work
is made available in the official python-sscha code as a
part of a new class called AiiDAEnsemble [16], which also
handles the automated submission of DFT jobs. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. On-the-fly active learning
The active learning workflow presented in this work
is made available in the official python-sscha code as a
part of a new class called AiiDAEnsemble [16], which also
handles the automated submission of DFT jobs. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. On-the-fly active learning
The active learning workflow presented in this work
is made available in the official python-sscha code as a
part of a new class called AiiDAEnsemble [16], which also
handles the automated submission of DFT jobs. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. On-the-fly active learning
The active learning workflow presented in this work
is made available in the official python-sscha code as a
part of a new class called AiiDAEnsemble [16], which also
handles the automated submission of DFT jobs. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. On-the-fly active learning
The active learning workflow presented in this work
is made available in the official python-sscha code as a
part of a new class called AiiDAEnsemble [16], which also
handles the automated submission of DFT jobs. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. On-the-fly active learning
The active learning workflow presented in this work
is made available in the official python-sscha code as a
part of a new class called AiiDAEnsemble [16], which also
handles the automated submission of DFT jobs. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. On-the-fly active learning
The active learning workflow presented in this work
is made available in the official python-sscha code as a
part of a new class called AiiDAEnsemble [16], which also
handles the automated submission of DFT jobs. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. On-the-fly active learning
The active learning workflow presented in this work
is made available in the official python-sscha code as a
part of a new class called AiiDAEnsemble [16], which also
handles the automated submission of DFT jobs. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. On-the-fly active learning
The active learning workflow presented in this work
is made available in the official python-sscha code as a
part of a new class called AiiDAEnsemble [16], which also
handles the automated submission of DFT jobs. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. On-the-fly active learning
The active learning workflow presented in this work
is made available in the official python-sscha code as a
part of a new class called AiiDAEnsemble [16], which also
handles the automated submission of DFT jobs. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. On-the-fly active learning
The active learning workflow presented in this work
is made available in the official python-sscha code as a
part of a new class called AiiDAEnsemble [16], which also
handles the automated submission of DFT jobs. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. On-the-fly active learning
The active learning workflow presented in this work
is made available in the official python-sscha code as a
part of a new class called AiiDAEnsemble [16], which also
handles the automated submission of DFT jobs. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. On-the-fly active learning
The active learning workflow presented in this work
is made available in the official python-sscha code as a
part of a new class called AiiDAEnsemble [16], which also
handles the automated submission of DFT jobs. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. On-the-fly active learning
The active learning workflow presented in this work
is made available in the official python-sscha code as a
part of a new class called AiiDAEnsemble [16], which also
handles the automated submission of DFT jobs. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. On-the-fly active learning
The active learning workflow presented in this work
is made available in the official python-sscha code as a
part of a new class called AiiDAEnsemble [16], which also
handles the automated submission of DFT jobs. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. On-the-fly active learning
The active learning workflow presented in this work
is made available in the official python-sscha code as a
part of a new class called AiiDAEnsemble [16], which also
handles the automated submission of DFT jobs. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. On-the-fly active learning
The active learning workflow presented in this work
is made available in the official python-sscha code as a
part of a new class called AiiDAEnsemble [16], which also
handles the automated submission of DFT jobs. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. On-the-fly active learning
The active learning workflow presented in this work
is made available in the official python-sscha code as a
part of a new class called AiiDAEnsemble [16], which also
handles the automated submission of DFT jobs. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. On-the-fly active learning
The active learning workflow presented in this work
is made available in the official python-sscha code as a
part of a new class called AiiDAEnsemble [16], which also
handles the automated submission of DFT jobs. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. On-the-fly active learning
The active learning workflow presented in this work
is made available in the official python-sscha code as a
part of a new class called AiiDAEnsemble [16], which also
handles the automated submission of DFT jobs. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. On-the-fly active learning
The active learning workflow presented in this work
is made available in the official python-sscha code as a
part of a new class called AiiDAEnsemble [16], which also
handles the automated submission of DFT jobs. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. On-the-fly active learning
The active learning workflow presented in this work
is made available in the official python-sscha code as a
part of a new class called AiiDAEnsemble [16], which also
handles the automated submission of DFT jobs. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. On-the-fly active learning
The active learning workflow presented in this work
is made available in the official python-sscha code as a
part of a new class called AiiDAEnsemble [16], which also
handles the automated submission of DFT jobs. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. 05 to call DFT and 0. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. 05 to call DFT and 0. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. 05 to call DFT and 0. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. 05 to call DFT and 0. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. 05 to call DFT and 0. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. 05 to call DFT and 0. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. 05 to call DFT and 0. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. 05 to call DFT and 0. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. 05 to call DFT and 0. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. 05 to call DFT and 0. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. 05 to call DFT and 0. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. 05 to call DFT and 0. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. 05 to call DFT and 0. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. 05 to call DFT and 0. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. 05 to call DFT and 0. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. 05 to call DFT and 0. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. 05 to call DFT and 0. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. 05 to call DFT and 0. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. 05 to call DFT and 0. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. 05 to call DFT and 0. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. 05 to call DFT and 0. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. 05 to call DFT and 0. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. 05 to call DFT and 0. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. 05 to call DFT and 0. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. 05 to call DFT and 0. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. 05 to call DFT and 0. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. 05 to call DFT and 0. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. 05 to call DFT and 0. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. 05 to call DFT and 0. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. 05 to call DFT and 0. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. 05 to call DFT and 0. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. 05 to call DFT and 0. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. 05 to call DFT and 0. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. 05 to call DFT and 0. The onthe-fly calculations used an (SGP) uncertainty threshold
of 0. 05 to call DFT and 0. 05 to call DFT and 0. 005 for the threshold to add the
sparse environments (corresponding to 10% of the uncertainty threshold). 05 to call DFT and 0. 005 for the threshold to add the
sparse environments (corresponding to 10% of the uncertainty threshold). 05 to call DFT and 0. 005 for the threshold to add the
sparse environments (corresponding to 10% of the uncertainty threshold). 05 to call DFT and 0. 005 for the threshold to add the
sparse environments (corresponding to 10% of the uncertainty threshold). 05 to call DFT and 0. 005 for the threshold to add the
sparse environments (corresponding to 10% of the uncertainty threshold). 05 to call DFT and 0. 005 for the threshold to add the
sparse environments (corresponding to 10% of the uncertainty threshold). 05 to call DFT and 0. 005 for the threshold to add the
sparse environments (corresponding to 10% of the uncertainty threshold). 05 to call DFT and 0. 005 for the threshold to add the
sparse environments (corresponding to 10% of the uncertainty threshold). 05 to call DFT and 0. 005 for the threshold to add the
sparse environments (corresponding to 10% of the uncertainty threshold). 05 to call DFT and 0. 005 for the threshold to add the
sparse environments (corresponding to 10% of the uncertainty threshold). 05 to call DFT and 0. 005 for the threshold to add the
sparse environments (corresponding to 10% of the uncertainty threshold). 05 to call DFT and 0. 005 for the threshold to add the
sparse environments (corresponding to 10% of the uncertainty threshold). 05 to call DFT and 0. 005 for the threshold to add the
sparse environments (corresponding to 10% of the uncertainty threshold). 05 to call DFT and 0. 005 for the threshold to add the
sparse environments (corresponding to 10% of the uncertainty threshold). 05 to call DFT and 0. 005 for the threshold to add the
sparse environments (corresponding to 10% of the uncertainty threshold). 05 to call DFT and 0. 005 for the threshold to add the
sparse environments (corresponding to 10% of the uncertainty threshold). 05 to call DFT and 0. 005 for the threshold to add the
sparse environments (corresponding to 10% of the uncertainty threshold). 05 to call DFT and 0. 005 for the threshold to add the
sparse environments (corresponding to 10% of the uncertainty threshold). 05 to call DFT and 0. 005 for the threshold to add the
sparse environments (corresponding to 10% of the uncertainty threshold). 05 to call DFT and 0. 005 for the threshold to add the
sparse environments (corresponding to 10% of the uncertainty threshold). 05 to call DFT and 0. 005 for the threshold to add the
sparse environments (corresponding to 10% of the uncertainty threshold). 05 to call DFT and 0. 005 for the threshold to add the
sparse environments (corresponding to 10% of the uncertainty threshold). 05 to call DFT and 0. 005 for the threshold to add the
sparse environments (corresponding to 10% of the uncertainty threshold). 05 to call DFT and 0. 005 for the threshold to add the
sparse environments (corresponding to 10% of the uncertainty threshold). 05 to call DFT and 0. 005 for the threshold to add the
sparse environments (corresponding to 10% of the uncertainty threshold). 05 to call DFT and 0. 005 for the threshold to add the
sparse environments (corresponding to 10% of the uncertainty threshold). 05 to call DFT and 0. 005 for the threshold to add the
sparse environments (corresponding to 10% of the uncertainty threshold). 05 to call DFT and 0. 005 for the threshold to add the
sparse environments (corresponding to 10% of the uncertainty threshold). 05 to call DFT and 0. 005 for the threshold to add the
sparse environments (corresponding to 10% of the uncertainty threshold). 05 to call DFT and 0. 005 for the threshold to add the
sparse environments (corresponding to 10% of the uncertainty threshold). 05 to call DFT and 0. 005 for the threshold to add the
sparse environments (corresponding to 10% of the uncertainty threshold). 05 to call DFT and 0. 005 for the threshold to add the
sparse environments (corresponding to 10% of the uncertainty threshold). 05 to call DFT and 0. 005 for the threshold to add the
sparse environments (corresponding to 10% of the uncertainty threshold). 05 to call DFT and 0. 005 for the threshold to add the
sparse environments (corresponding to 10% of the uncertainty threshold). 05 to call DFT and 0. 005 for the threshold to add the
sparse environments (corresponding to 10% of the uncertainty threshold). The active-learning simulations are carried out over the
entire temperature range shown in Fig. 2, with a temperature step of ∆T = 50 K, and using the “optimal-guess”
strategy (see Optimal guess strategy section in Methods). The active-learning simulations are carried out over the
entire temperature range shown in Fig. 2, with a temperature step of ∆T = 50 K, and using the “optimal-guess”
strategy (see Optimal guess strategy section in Methods). The active-learning simulations are carried out over the
entire temperature range shown in Fig. 2, with a temperature step of ∆T = 50 K, and using the “optimal-guess”
strategy (see Optimal guess strategy section in Methods). The active-learning simulations are carried out over the
entire temperature range shown in Fig. 2, with a temperature step of ∆T = 50 K, and using the “optimal-guess”
strategy (see Optimal guess strategy section in Methods). The active-learning simulations are carried out over the
entire temperature range shown in Fig. 2, with a temperature step of ∆T = 50 K, and using the “optimal-guess”
strategy (see Optimal guess strategy section in Methods). The active-learning simulations are carried out over the
entire temperature range shown in Fig. 2, with a temperature step of ∆T = 50 K, and using the “optimal-guess”
strategy (see Optimal guess strategy section in Methods). The active-learning simulations are carried out over the
entire temperature range shown in Fig. 2, with a temperature step of ∆T = 50 K, and using the “optimal-guess”
strategy (see Optimal guess strategy section in Methods). The active-learning simulations are carried out over the
entire temperature range shown in Fig. 2, with a temperature step of ∆T = 50 K, and using the “optimal-guess”
strategy (see Optimal guess strategy section in Methods). The active-learning simulations are carried out over the
entire temperature range shown in Fig. 2, with a temperature step of ∆T = 50 K, and using the “optimal-guess”
strategy (see Optimal guess strategy section in Methods). The active-learning simulations are carried out over the
entire temperature range shown in Fig. 2, with a temperature step of ∆T = 50 K, and using the “optimal-guess”
strategy (see Optimal guess strategy section in Methods). The active-learning simulations are carried out over the
entire temperature range shown in Fig. 2, with a temperature step of ∆T = 50 K, and using the “optimal-guess”
strategy (see Optimal guess strategy section in Methods). The active-learning simulations are carried out over the
entire temperature range shown in Fig. 2, with a temperature step of ∆T = 50 K, and using the “optimal-guess”
strategy (see Optimal guess strategy section in Methods). The active-learning simulations are carried out over the
entire temperature range shown in Fig. 2, with a temperature step of ∆T = 50 K, and using the “optimal-guess”
strategy (see Optimal guess strategy section in Methods). The active-learning simulations are carried out over the
entire temperature range shown in Fig. 2, with a temperature step of ∆T = 50 K, and using the “optimal-guess”
strategy (see Optimal guess strategy section in Methods). The active-learning simulations are carried out over the
entire temperature range shown in Fig. 2, with a temperature step of ∆T = 50 K, and using the “optimal-guess”
strategy (see Optimal guess strategy section in Methods). The active-learning simulations are carried out over the
entire temperature range shown in Fig. 2, with a temperature step of ∆T = 50 K, and using the “optimal-guess”
strategy (see Optimal guess strategy section in Methods). The active-learning simulations are carried out over the
entire temperature range shown in Fig. 2, with a temperature step of ∆T = 50 K, and using the “optimal-guess”
strategy (see Optimal guess strategy section in Methods). The active-learning simulations are carried out over the
entire temperature range shown in Fig. 2, with a temperature step of ∆T = 50 K, and using the “optimal-guess”
strategy (see Optimal guess strategy section in Methods). The active-learning simulations are carried out over the
entire temperature range shown in Fig. 2, with a temperature step of ∆T = 50 K, and using the “optimal-guess”
strategy (see Optimal guess strategy section in Methods). The active-learning simulations are carried out over the
entire temperature range shown in Fig. 2, with a temperature step of ∆T = 50 K, and using the “optimal-guess”
strategy (see Optimal guess strategy section in Methods). The active-learning simulations are carried out over the
entire temperature range shown in Fig. 2, with a temperature step of ∆T = 50 K, and using the “optimal-guess”
strategy (see Optimal guess strategy section in Methods). The active-learning simulations are carried out over the
entire temperature range shown in Fig. 2, with a temperature step of ∆T = 50 K, and using the “optimal-guess”
strategy (see Optimal guess strategy section in Methods). The active-learning simulations are carried out over the
entire temperature range shown in Fig. 2, with a temperature step of ∆T = 50 K, and using the “optimal-guess”
strategy (see Optimal guess strategy section in Methods). The active-learning simulations are carried out over the
entire temperature range shown in Fig. 2, with a temperature step of ∆T = 50 K, and using the “optimal-guess”
strategy (see Optimal guess strategy section in Methods). The active-learning simulations are carried out over the
entire temperature range shown in Fig. 2, with a temperature step of ∆T = 50 K, and using the “optimal-guess”
strategy (see Optimal guess strategy section in Methods). The active-learning simulations are carried out over the
entire temperature range shown in Fig. 2, with a temperature step of ∆T = 50 K, and using the “optimal-guess”
strategy (see Optimal guess strategy section in Methods). The active-learning simulations are carried out over the
entire temperature range shown in Fig. 2, with a temperature step of ∆T = 50 K, and using the “optimal-guess”
strategy (see Optimal guess strategy section in Methods). The active-learning simulations are carried out over the
entire temperature range shown in Fig. 2, with a temperature step of ∆T = 50 K, and using the “optimal-guess”
strategy (see Optimal guess strategy section in Methods). The active-learning simulations are carried out over the
entire temperature range shown in Fig. 2, with a temperature step of ∆T = 50 K, and using the “optimal-guess”
strategy (see Optimal guess strategy section in Methods). The active-learning simulations are carried out over the
entire temperature range shown in Fig. 2, with a temperature step of ∆T = 50 K, and using the “optimal-guess”
strategy (see Optimal guess strategy section in Methods). The active-learning simulations are carried out over the
entire temperature range shown in Fig. 2, with a temperature step of ∆T = 50 K, and using the “optimal-guess”
strategy (see Optimal guess strategy section in Methods). The active-learning simulations are carried out over the
entire temperature range shown in Fig. 2, with a temperature step of ∆T = 50 K, and using the “optimal-guess”
strategy (see Optimal guess strategy section in Methods). The active-learning simulations are carried out over the
entire temperature range shown in Fig. 2, with a temperature step of ∆T = 50 K, and using the “optimal-guess”
strategy (see Optimal guess strategy section in Methods). The active-learning simulations are carried out over the
entire temperature range shown in Fig. 2, with a temperature step of ∆T = 50 K, and using the “optimal-guess”
strategy (see Optimal guess strategy section in Methods). The active-learning simulations are carried out over the
entire temperature range shown in Fig. 2, with a temperature step of ∆T = 50 K, and using the “optimal-guess”
strategy (see Optimal guess strategy section in Methods).Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
On-the-fly active learning scheme combining SSCHA method with Bayesian ML potentials can greatly accelerate first-principles accurate finite temperature calculations, enabling efficient application to a wide range of temperatures and accurate capture of challenging phase transitions.
-
The improved system can lower the SSCHA first-principles workload by 98.4 – 99.8%, as demonstrated for Li2O and CsPbI3, reducing DFT calls significantly compared to previous methods like MD-based active-learning schemes.
-
The framework allows for targeted exploration of the potential energy surface, proving that
the relevant physics is contained in only a few configurations of the phase space
when targeting specific physical properties such as thermodynamic free energies. -
The system can accurately predict the phase transition temperature governing the (exceptional) optoelectronic properties of CsPbI3 by using SSCHA to find the Gibbs free energy difference between competing phases, which corresponds to
the temperature T for which delta G delta alpha equals zero.
-
The improved system can calculate linear thermal expansion coefficients with high accuracy, such as a value of 2.8 x 10-5 K-1 at T0 = 1000 K for Li2O, matching experimental values within a tight threshold.
Abstract
Materials underpin modern technologies, from energy harvesting, storage, and conversion to information and communication technologies. Their functionality is often governed by the interplay between competing phases, and thermodynamic behavior shapes microscopic properties and ultimately determines technological performance; for instance, the light absorption of inorganic metal-halide perovskites in solar cells. Accurately predicting crystal thermodynamics, however, remains a major challenge for computational approaches because strong anharmonic effects require extensive sampling of the potential energy surface. Here, we present an on-the-fly Bayesian framework, combined with the stochastic self-consistent harmonic approximation, for learning first-principles potentials. This approach enables the prediction of thermodynamic properties over a broad temperature range with first-principles accuracy while requiring training on only a few atomic configurations. A relevant finding is that the majority of the required structures are obtained from a single inexpensive dataset generation step, demonstrating that the relevant physics for accurately predicting solid-state thermodynamics is encoded in a few physically meaningful configurations. We investigate the thermodynamic and dynamical properties of Li 2 O, α --CsPbI 3, and δ --CsPbI 3, requiring only 44, 256, and 50 total-energy calculations, respectively. Notably, we show that this framework accurately captures the extreme thermal expansion of, and the phase diagram of CsPbI 3 with a predicted transition temperature within 20 K of experiment. More broadly, the method presented opens an interesting route toward accelerated materials engineering under realistic conditions for a wide range of technologically relevant applications, including batteries, photovoltaics, and memristors.
Related papers
- Structural contribution to light-induced gap suppression in Ta 2 NiSe 5
- Anatomy of fast current-induced skyrmion motion in synthetic antiferromagnets
- Magnetic control of an exciton-polariton condensate in a van der Waals magnet
- High-pressure elastic properties of GeO2 polymorphs up to 120 GPa
- Decoupling Strain-Rate Sensitivity and Deformation Length Scale Effects in Neutron-Irradiated Tungsten: A Coupled Nano-Indentation, HR-EBSD and Crystal Plasticity Study
- Magnetism modulation in Co3Sn2S2 by current-assisted domain wall motion