A General-Purpose and Efficient Machine-Learned Potential for SiC from Ambient to Extreme Environments

arXiv:2510.01827 · cond-mat.mtrl-sci · Submitted 2025-10-02 · Read on arXiv

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: We're looking at the paper: "A General-Purpose and Efficient Machine-Learned Potential for SiC from Ambient to Extreme Environments".

Mira: Here's the gist of the paper. A computationally efficient, general-purpose machine-learned interatomic potential (ML-IAP) named tabGAP is developed for SiC systems, capable of performing multimillion-atom molecular dynamics simulations over microsecond timescales.

Kai: tabGAP is designed to map a broad pressure-temperature phase diagram and threshold displacement energy (TDE) distributions for both the 2H and 3C polymorphs across conditions ranging from ambient states up to extreme high-pressure/high-temperature regimes and high-energy cascade damage.

Mira: The method achieves a favorable balance between accuracy, robustness, transferability, and computational cost compared to other empirical and ML-IAPs tested.

Kai: The training database for tabGAP consists of 3460 configurations encompassing 185,542 local atomic environments sampled from diverse domains: dispersed configurations (dimers and trimers), crystalline SiC polymorphs (2H, 3C, 4H, 6H, RS), pure elemental phases (Si and C), and disordered bulk phases (amorphous and molten states).

Paper summary: Kai: We're starting with how tabGAP maps out this broad pressure temperature phase diagram for silicon carbide, showing transitions across various polymorphs like 2H and 3C when you push the conditions really hard. The authors mention they used free energy simulations and direct large scale molecular dynamics simulations of melting and decomposition to build this map.

Mira: They compare this map against experimental data and density functional theory calculations, which reveals transitions such as the 3C to 6H transition occurring around two thousand kelvin at eighty mega pascals. That is a specific transition point they found.

Lev: From a computational overhead view, the construction of this P-T phase diagram involved liquid vacuum molecular dynamics simulations to define the gas phase boundary and two phase npt simulations to validate coexistence boundaries between different silicon carbide phases.

Kai: They also used solubility simulations to cross check the liquid silicon plus carbon graphitic minus liquid silicon carbide boundary, which suggests that silicon carbide will pass through a metastable decomposition phase and a solid carbon liquid silicon mixing phase during heating and melting at normal to high pressures.

Mira: And they noted that the decomposed carbon structure depends on pressure, suggesting graphite at low pressure and diamond at high pressure for that decomposed carbon part. This shows how the environment directly affects the resulting material state.

Lev: The method's computational speed is roughly one order of magnitude slower than tersoff zbl because it uses a larger cutoff radius of five angstroms instead of three angstroms, but they say this trade off is justified by its broad coverage and robustness for targeted applications.

Conclusion: Kai: So this paper, titled "A General-Purpose and Efficient Machine-Learned Potential for SiC from Ambient to Extreme Environments," focuses on tabGAP, which is this machine learned interatomic potential for silicon carbide that's supposed to cover everything from normal conditions all the way up to really high pressure and heat.

Mira: Right, it’s a general purpose model, so they’re trying to make one thing that works for a huge range of stuff instead of having ten different models for every tiny case.

Lev: From an engineering standpoint, that broad coverage is good because you don't have to rewrite the software when you move from ambient room temperature to something hot and under immense pressure.

Kai: The authors say they trained this model on a database with a lot of configurations, including everything from simple pairs of atoms up to complex crystalline structures like 2H and 3C polytypes #pg1.

Mira: They’re using density functional theory data for training, spanning temperatures up to six thousand kelvin and pressures up to one hundred one gigapascals. That is a massive range of conditions they are mapping out.

Lev: The validation showed that the energies and forces this potential generates are pretty consistent with those dft calculations, with an error of about zero point thirteen electron volts per atom for bulk silicon carbide.

Kai: It’s a solid result for accuracy across that wide spectrum, but they did admit there are some minor discrepancies in things like the optical branches and the rs structure because they had to make some trade offs during the fitting process.

Mira: So while it’s very good for mapping out phase diagrams and general behavior, it’s not perfect for super fine details like how many vacancies form or how defects evolve over a very long time without more specific testing.

Lev: That makes sense; you can't expect a potential that handles melting at high pressure to be the absolute best model for tracking tiny atomic rearrangements over millions of years.

Kai: The main implication is that this tool allows researchers to do large scale molecular dynamics simulations on silicon carbide systems under conditions we couldn't realistically access before.

Mira: It provides a practical way to explore phase transitions, like seeing exactly when the 3C structure transforms into the 6H structure around two thousand kelvin and eighty mega pascals #pg3.

Lev: It’s a powerful tool for materials science now, but it still requires careful application because you have to keep those limitations in mind when you use it for things like defect statistics.

Jintong Wu, Zhuang Shao, Junlei Zhao, Flyura Djurabekova, Kai Nordlund, Fredric Granberg, Qingmin Zhang

Hong Kong Microelectronics Research and Development Institute of Hong Kong, Department of Physics at University of Helsinki, School of Energy and Power Engineering at Xi’an Jiaotong University, Helsinki Institute of Physics at University of Helsinki

cond-mat.mtrl-sci

Submitted: 2025-10-02

Updated: 2026-11-14

Comments: Main text: 16 pages, 7 figures; Supplementary Information: 51 pages, 36 figures

Code: https://github.com/libAtoms/QUIP

License: http://creativecommons.org/licenses/by-nc-nd/4.0/

Importance score: 77/100

The gist: A computationally efficient, general-purpose machine-learned interatomic potential (ML-IAP) named tabGAP is developed for SiC systems…

Key concepts

Machine-Learned Interatomic Potential (ML-IAP)
A computational method that uses machine learning to create an accurate mathematical function describing the interatomic forces in a material. This allows researchers to simulate how atoms interact over long times without needing expensive, high-level quantum calculations for every step.
Phase Diagram Mapping
The ability of the potential to predict stable phases of SiC under varying pressure and temperature conditions. tabGAP was used to construct this map by running simulations that mimic real-world extreme environments, helping engineers understand when SiC will melt or transform into different structures.
Threshold Displacement Energy (TDE)
A measure derived from the potential that quantifies the energy required for an atom to move a certain distance within the material. This concept helps predict how easily defects or atoms can be displaced under stress, which is crucial for understanding material stability and damage under high-energy events.

Terminology

Summary

A computationally efficient, general-purpose machine-learned interatomic potential (ML-IAP) named tabGAP is developed for SiC systems, capable of performing multimillion-atom molecular dynamics simulations over microsecond timescales. p. 0

Where the paper says it

The rapid advancement of SiC-based applications has been complemented by computational modeling studies, including both ab initio and classical atomistic approaches. In this work, we develop a computationally efficient, general-purpose machine-learned interatomic potential (ML-IAP) capable of multimillion-atom molecular dynamics simulations over microsecond timescales.

tabGAP is designed to map a broad pressure-temperature phase diagram and threshold displacement energy (TDE) distributions for both the 2H and 3C polymorphs across conditions ranging from ambient states up to extreme high-pressure/high-temperature regimes and high-energy cascade damage. p. 0
Where the paper says it

Using the ML-IAP, we map a broad pressure-temperature phase diagram and the threshold displacement energy distributions for the 2H and 3C polymorphs. Across a benchmark covering conditions from ambient to extreme, including high-pressure/hightemperature states and high-energy cascade damage, tabGAP provides a favorable balance of accuracy, robustness, transferability, and computational cost among the tested empirical and ML-IAPs.

The method achieves a favorable balance between accuracy, robustness, transferability, and computational cost compared to other empirical and ML-IAPs tested. p. 0
Where the paper says it

Using the ML-IAP, we map a broad pressure-temperature phase diagram and the threshold displacement energy distributions for the 2H and 3C polymorphs. Across a benchmark covering conditions from ambient to extreme, including high-pressure/hightemperature states and high-energy cascade damage, tabGAP provides a favorable balance of accuracy, robustness, transferability, and computational cost among the tested empirical and ML-IAPs.

The training database for tabGAP consists of 3460 configurations encompassing 185,542 local atomic environments sampled from diverse domains: dispersed configurations (dimers and trimers), crystalline SiC polymorphs (2H, 3C, 4H, 6H, RS), pure elemental phases (Si and C), and disordered bulk phases (amorphous and molten states). p. 1

Where the paper says it

(iii) Pure elemental phases: To accurately simulate incongruent melting processes under high temperatures and pressures, we included diverse configurations of pure carbon (diamond, graphite, graphene, amorphous carbon) and silicon (diamond structure, molten, amorphous). (iv) Disordered bulk phases: amorphous and molten states were generated by rapid heating of 3C-SiC to 3000 K in AIMD with NV T ensembles until complete melting, followed by rapid quenching (cooling rate of 1014 K/s).

The training data is derived entirely from Density Functional Theory (DFT) reference data spanning an extensive range of temperatures up to 6500 K and pressures up to 110 GPa. p. 1
Where the paper says it

Notably, while the tabGAP employs simple, low-dimensional descriptors similar to those of empirical IAPs, it is fundamentally a ML-IAP: its ∼2000 parameters (weights) are fitted via sparse Gaussian process regression without assuming a fixed analytical form, and its training relies entirely on DFT reference data rather than experimental fitting. Trained on DFT data covering an extreme range of temperatures and pressures (up to 6500 K, 110 GPa), our potential reproduces selected DFT reference data and validation properties with useful accuracy while retaining high computational efficiency.

Validation confirms that the energies and forces generated by tabGAP are broadly consistent with DFT calculations over wide ranges, exhibiting aggregate root mean square errors (RMSE) of 0. p. 2
Where the paper says it

Most data collapse around the diagonal in Fig. 1c and 1d, with aggregate root mean square errors (RMSE) of 0.

013 eV/atom and 0. p. 2
Where the paper says it

1c and 1d, with aggregate root mean square errors (RMSE) of 0. 013 eV/atom and 0.

679 eV/˚A for bulk SiC, supporting its utility in studying thermodynamic responses under extreme conditions. p. 2
Where the paper says it

679 eV/˚A. This agreement over a wide range supports, but does not by itself prove, the use of tabGAP for studying the thermodynamic response of SiC under extreme temperatures and pressures.

While the potential reproduces DFT low-energy polytype groups at zero pressure and temperature, minor discrepancies exist in the optical branches and RS structure due to the efficiency compromises made during fitting. p. 3
Where the paper says it

2 are seen for the optical branches and the RS structure. These discrepancies reflect the compromise made in fitting a highly efficient potential over a broad energy and pressure range.

The tabGAP method is employed to construct a P-T phase diagram for SiC through large numbers of free-energy simulations and direct large-scale MD simulations of melting and decomposition. p. 5

Where the paper says it

Second, it provides useful guidance for SiC phase engineering, as experimentally mapping the P-T phase diagram is extremely challenging. Therefore, we performed a large number of freeenergy simulations and direct large-scale MD simulations of SiC melting and decomposition to trace the phase boundaries and construct a broad tabGAP/PBE-level P-T phase map (see details in the Methods and Supplementary Note 9), as shown in Fig.

This map is compared against experimental data and DFT calculations, providing guidance for SiC phase engineering, revealing transitions such as the 3C to 6H transition occurring around 2000 K at 80 MPa. p. 5
Where the paper says it

Therefore, we performed a large number of freeenergy simulations and direct large-scale MD simulations of SiC melting and decomposition to trace the phase boundaries and construct a broad tabGAP/PBE-level P-T phase map (see details in the Methods and Supplementary Note 9), as shown in Fig. The P-T phase diagram predicted by tabGAP is compared to available experimental data and DFT calculations.

Furthermore, the model was tested for radiation damage simulations and TDE trends in the 3C-SiC environment, showing close consistency with DFT and TGAP regarding defect statistics and cascade morphologies, although Tersoff/ZBL predicted higher TDEs for Si PKAs. p. 9
Where the paper says it

Furthermore, tabGAP captures several radiation-response trends in SiC, such as TDE and QSD, while retaining important limitations in some defect-related quantities. Although TGAP gives somewhat better agreement for defect formation energies, both models exhibit close consistency in defect statistics and collision cascade morphologies for 3C-SiC, as validated in Supplementary Note 10.

When benchmarked against other ML-IAPs (TGAP [58], FLARE [75], UF3 [55]) and empirical IAPs (Tersoff/ZBL [42], Vashishta [44]), tabGAP demonstrated the most balanced aggregate performance across a wide spectrum of material properties and computational efficiencies, being comparable to classical modified embedded atom method potentials on GPUs. p. 9

Where the paper says it

Most existing IAPs are designed for a single SiC polytype and developed for specific properties within particular application contexts. To systematically evaluate the developed tabGAP for SiC, we benchmarked its performance against currently available ML-IAPs (TGAP [58], FLARE [75], UF3 [55]) and widely used empirical IAPs (Tersoff/ZBL [42], Vashishta [44]).

The authors conclude that tabGAP is an efficient, general-purpose tool for large-scale SiC simulations under extreme conditions involving high temperatures, high pressures, and irradiation. p. 1
Where the paper says it

Trained on DFT data covering an extreme range of temperatures and pressures (up to 6500 K, 110 GPa), our potential reproduces selected DFT reference data and validation properties with useful accuracy while retaining high computational efficiency. Its transferability across ambient and extreme conditions allows us to compare with available experimental observations and to construct a detailed tabGAP/PBE-level pressure-temperature phase map of SiC, together with TDE maps and million-atom simulations of large-scale collision cascades for different polymorphs.

However, it is not recommended as a high-precision model for vacancy formation energies or long-time defect evolution without further dedicated validation. p. 9
Where the paper says it

In contrast, the Tersoff/ZBL potential systematically predicted significantly higher levels of disorder and visibly different cascade morphologies, tending to form larger atomic clusters while tabGAP and TGAP left defects more dispersed and isolated. This combined evaluation supports the use of tabGAP for primary high-energy cascade simulations, but it does not establish uniformly quantitative accuracy for isolated vacancy thermodynamics or long-time defect evolution.

The construction of the P-T phase diagram also involved liquid-vacuum MD simulations to define the gas-phase boundary and two-phase NPT simulations to validate coexistence boundaries between different SiC phases. p. 6

Where the paper says it

7 LSi + Cdia + 3 C 10 RS 3C LSi + Cdia LSi + Cgra LSiC LSi+C Gas DFT CALPHY Exp NPT coexistence NVT liquid-vacuum Exp (3C-6H) Exp (6H-3C) CALPHY (6H-3C) CALPHY (3C-6H) Solubility NPH i ii iii iv v vi vii viii @200GPa, 8000K Si C i ii iii iv v vi vii viii ix b ix a FIG. Pressure-temperature phase diagram of SiC.

Solubility simulations were used to cross-check the LSi + Cgra–LSiC boundary, suggesting that SiC will pass through a metastable decomposition phase and a SolidC-LiquidSi mixing phase during heating and melting at normal to high pressures, with the decomposed carbon structure depending on pressure (graphite at low pressure, diamond at high pressure). p. 7
Where the paper says it

Our results predict that SiC will pass through a metastable decomposition phase and the SolidC-LiquidSi mixing phase in the process of heating and melting at normal to high pressures. The structure of the decomposed C clusters depends on pressure; graphite at low pressure and diamond at high pressure.

The computational speed is roughly one order of magnitude slower than Tersoff/ZBL due to a larger cutoff radius (5 Å vs. p. 9
Where the paper says it

7, the performance is assessed across a broad spectrum of material properties and computational efficiencies, where a larger enclosed area indicates a more favorable aggregate score within the author-defined metric set. The evaluation metrics include: (i) crystalline phases, including phase ordering and phase diagram prediction; (ii) key structural descriptors of the amorphous phase (RDF, bond angle distribution, sp2/sp3 ratio); (iii) mechanical and defect properties (elastic constants, defect formation energies Ef); (iv) response to radiation damage (QSD and TDE); (v) computational speed denoted by ηCPU and ηGPU for CPU and GPU architectures, respectively.

3 Å), but this trade-off is justified by its broad coverage and robustness for targeted applications. p. 0
Where the paper says it

Using the ML-IAP, we map a broad pressure-temperature phase diagram and the threshold displacement energy distributions for the 2H and 3C polymorphs. Across a benchmark covering conditions from ambient to extreme, including high-pressure/hightemperature states and high-energy cascade damage, tabGAP provides a favorable balance of accuracy, robustness, transferability, and computational cost among the tested empirical and ML-IAPs.

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

  1. Develop a general-purpose machine-learned interatomic potential (ML-IAP) capable of multimillion-atom molecular dynamics simulations over microsecond timescales for SiC across ambient to extreme environments, as demonstrated by tabGAP. This allows the improved system to map a broad pressure-temperature phase diagram and threshold displacement energy distributions for 2H and 3C polymorphs.

  2. Implement a robust method for constructing pressure-temperature phase diagrams by performing large numbers of freeenergy simulations and direct large-scale MD simulations, providing guidance for SiC phase engineering where experimentally mapping is challenging.

  3. Enhance the system's capability to predict phase transition behavior by using the tabGAP/PBE level P-T map to compare predicted boundaries with experimental data and DFT calculations, accounting for the competition between thermodynamic stability and kinetic stability.

  4. Improve accuracy in modeling radiation damage by using quasi-static drag calculations and comparing results directly to DFT and TGAP ML-IAPs, allowing the system to predict Threshold Displacement Energy (TDE) trends.

  5. Enable detailed study of structural transformations in disordered phases by using tabGAP to compute phonon dispersions for three polymorphs, capturing near-equilibrium vibrational properties with reasonable fidelity while noting limitations regarding long-range electrostatics.

  6. Refine phase transition modeling under extreme conditions by simulating the decomposition process in high pressure regions, predicting atomistic details such as the carbon cluster structure (graphite at low pressure and diamond at high pressure).

  7. Improve accuracy in describing short-range many-body behavior during cascade simulations by using tabGAP to predict TDE values for primary knock-on atoms, noting that In 3C the difference between Si and C TDEs in 3C is about twice as large with Tersoff/ZBL as with tabGAP.

  8. Provide a balanced assessment of various IAPs by benchmarking them against others using a multi-dimensional radar chart, allowing for selection of potentials like tabGAP that offer the most balanced aggregate performance among the tested IAPs when considering all metrics together.

  9. Offer quantitative context for equilibrium structural and stiffness metrics by providing component-wise elastic results, noting that deviations in lattice constants and bulk moduli are below approximately 0.6% and about 1.4–5.0%.

  10. Improve the system's ability to predict the solubility boundary between LSi + Cgra and LSiC by performing MD solubility simulations, utilizing a Van’t Hoff-type relation to determine the temperature T star(P) at which xC equals 0.5.

Abstract

Silicon carbide (SiC) polymorphs are widely employed as nuclear materials, mechanical components, and wide-bandgap semiconductors. The rapid advancement of SiC-based applications has been complemented by computational modeling studies, including both ab initio and classical atomistic approaches. In this work, we develop a computationally efficient, general-purpose machine-learned interatomic potential (ML-IAP) capable of multimillion-atom molecular dynamics simulations. Building on the tabulated Gaussian approximation potential (tabGAP) framework, we introduce an additional embedding term that captures local atomic coordination, improving accuracy and generality, while retaining high computational efficiency. Using this ML-tabGAP model, we map a broad pressure-temperature phase diagram, simulate the threshold displacement energy distributions for the 2H and 3C polymorphs, and simulate primary irradiation damage in 3C. Across a benchmark covering conditions from ambient to extreme, including high-pressure/high-temperature states and high-energy cascade damage, tabGAP provides a favorable balance of accuracy, robustness, transferability, and computational cost among the tested empirical and ML-IAPs.

Sources

Related papers