A Differentiable DFT-Based Framework for Inverse Materials Design

arXiv:2606.25502 · cond-mat.mtrl-sci, physics.comp-ph · Submitted 2026-06-24 · 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: I'm Kai.

Mira: And I'm Mira, here with Lev.

Kai: We're looking at the paper: "A Differentiable DFT-Based Framework for Inverse Materials Design".

Mira: Here's the gist of the paper. Discovering solid-state materials with target properties remains a central challenge in computational materials science<ref:2606.25502#pg2> Here we develop a first-principles inverse-design framework, integrating reverse-mode automatic differentiation into KKR-CPA—the Korringa–Kohn–Rostoker method with the coherent potential approximation—where atomic compositions are continuous variables to be optimized<ref:2606.25502#pg2> Reverse-mode AD yields gradients of objective functions with respect to composition at a cost independent of the number of candidate elements, enabling gradient-based optimization to identify materials from compositional spaces spanning dozens of elements<ref:2606.25502#pg2> In this framework, any computable quantity can serve as the objective<ref:2606.25502#pg2> We demonstrate this generality through two contrasting applications, magnetic alloys and halfmetals, yielding candidates such as (Lu0.553Yb0.447)(Co0.759Fe0.241)2Fe3 and FeZr(Sb0.94Te0.06)<ref:2606.25502#pg2> Our framework offers a physically grounded route from a target property to the material that realizes it<ref:2606.25502#pg2> Framework Foundation The core of the method is integrating reverse-mode AD into KKR-CPA, treating atomic compositions as continuous variables for optimization<ref:2606.25502#pg2> The KKR-CPA method is a first-principles framework in which compositions are intrinsically continuous and differentiable<ref:2606.25502#pg2> This approach requires no supercell, and its cost is nearly independent of the number of elements mixed on each site<ref:2606.25502#pg2> The CPA performs the configurational average over substitutional disorder analytically, assigning each site continuous concentrations x(s)i specifying the probability that element i occupies site s<ref:2606.25502#pg2> The total magnetic moment is defined as the integrated spin density µtot(x) = Z

ρ↑(r; x) − ρ↓(r; x): dr<ref:2606.25502#pg2> Optimization Mechanism Gradient-based inverse design follows the gradient of an objective straight toward a target, avoiding both the brute-force cost of screening and the training-data dependence of data-driven methods<ref:2606.25502#pg2> Automatic differentiation computes exact gradients with respect to many variables for essentially any differentiable objective, and now underpins gradient-based inverse design across fields<ref:2606.25502#pg2> Reverse mode AD computes the gradient of a scalar output with respect to all input parameters in a single backward pass, at a cost independent of the parameter count<ref:2606.25502#pg2> The workflow implements this as an iterative loop: KKR-CPA forward evaluation produces O(x), reverse-mode AD provides ∇xO, and the composition is updated<ref:2606.25502#pg2> Generality and Application The framework proves its generality by optimizing contrasting targets, such as the total magnetic moment, µtot<ref:2606.25502#pg2> It also optimizes a target based on the shape of the density of states (DOS) to realize a half-metal<ref:2606.25502#pg2> For example, in designing a half-Heusler XY Z compound, the optimization converges to FeZr(Sb0.94Te0.06), which is close to the known half-metal FeZrSb<ref:2606.25502#pg2> The framework can optimize for any target that KKR-CPA can compute, including electronic structure of disordered systems, magnetism and Curie temperatures<ref:2606.25502#pg2> Scalability and Robustness The method scales to inverse design in large compositional spaces because the gradient cost is independent of the parameter count<ref:2606.25502#pg2> For a 30-element alloy, gradient-based optimization drives unneeded elements toward zero, selecting both the elements and their proportions<ref:2606.25502#pg2> The per-iteration cost increased by only a factor of 3 when moving from 3 to 30 compositional parameters<ref:2606.25502#pg2> Furthermore, the framework can search for dopants that most improve an existing functional material by tracing continuous substitution pathways from a known compound<ref:2606.25502#pg2> Future Directions The present implementation uses the atomic-sphere approximation, but future work can incorporate a full-potential treatment to improve accuracy<ref:2606.25502#pg2> Thermodynamic stability and lattice relaxation are not yet addressed, but these limitations can be lifted in future work by differentiating the total energy with respect to atomic positions<ref:2606.25502#pg2> The same differentiability extends beyond atomic positions to external parameters such as applied magnetic and electric fields, pointing toward a framework that inverse-designs composition, lattice, and external fields together<ref:2606.

Title and authors: Kai: Let's talk about this paper, "A Differentiable DFT-Based Framework for Inverse Materials Design," and the authors who wrote it. The title itself tells us they are aiming to bridge the gap between a target property and the actual material composition.

Mira: The authors are from institutions like the Institute for Solid State Physics at University of Tokyo and Calmarion Inc., which shows this is coming from a place with strong materials science and experimental links.

Lev: From an engineering standpoint, what I find interesting is that they are focusing on how to make this differentiable so we can use gradient descent, which means we are looking for the slope toward our target property.

Kai: It’s not just about having a nice mathematical framework; it’s about using reverse mode automatic differentiation to compute those gradients with respect to composition at a cost that is independent of the number of candidate elements.

Mira: That cost independence is what makes it feasible for exploring compositional spaces that have dozens of elements, something traditional high-throughput screening simply can't do efficiently.

Lev: If we were trying to run this on real hardware, the overhead would land mostly in the forward evaluation cost of KKR-CPA, but since the gradient calculation itself is cheap relative to the number of variables, that helps a lot with iteration speed.

Kai: So, we're looking at this paper because it offers a path that avoids both the brute force cost of screening and the training data dependence you see in some AI inverse design methods.

Mira: The summary explains that this approach allows us to take any computable quantity—like total magnetic moment—as the objective function, giving us a physically grounded route from a target property to the material.

Lev: That physical grounding is what matters for me; it means we are not just optimizing numbers, we are optimizing things that actually have physical meaning in terms of magnetism or electronic structure.

Kai: And they demonstrate this generality by applying it to magnetic alloys and half metals, showing that the framework isn't just a niche tool but can apply across different material classes.

Mira: It’s interesting how they show this capability by optimizing targets based on things like the shape of the density of states to achieve a specific electronic feature, like being a half metal.

Lev: That specific result they mention, FeZr(Sb0.94Te0.06) as an example for the half-Heusler compound, shows that this method can converge on materials that are close to known functional materials <ref:2606.25502#pg1>.

The paper's summary: Kai: So let's get into what the paper actually says about how this framework works in practice. They start by laying out the problem: finding a material for a target property is hard because the space of candidate materials is enormous and you can’t just test them all one by one.

Mira: The summary explains that they address this by replacing the discrete choice of elements on lattice sites with continuous concentrations, using KKR-CPA. This replaces supercell methods or virtual crystal approximations where you have to choose a specific element at each site.

Lev: The paper explains that KKR-CPA performs an analytical configurational average over substitutional disorder, meaning each site has a continuous concentration x i specifying the probability of element i occupying that site.

Kai: This is important because it means the method doesn't need a supercell and its cost isn't dependent on the number of elements mixed on each site, which is a huge win for scaling up.

Mira: Then they introduce reverse mode automatic differentiation, which lets us compute gradients of our objective function with respect to these composition variables at a cost independent of the parameter count.

Lev: So the workflow is an iterative loop: you do a forward evaluation in KKR-CPA to get the quantity, then you use reverse mode AD to get the gradient, and then you update your composition variables.

Kai: The summary emphasizes that this allows for gradient-based optimization to find materials from compositional spaces spanning dozens of elements by following the direction of steepest descent toward our target property.

Mira: They show that any computable quantity can serve as the objective function, which gives us a lot of flexibility in choosing what we want to optimize for in the first place.

Lev: That generality is impressive, but we have to remember that this relies on KKR-CPA being a first principles framework where compositions are intrinsically continuous and differentiable.

Kai: It’s about making sure the underlying physics—the DFT calculations—are set up in a way that allows for this differentiability, which is the foundation of the whole framework.

The paper's improvements: Mira: Now, let's talk about what they suggest as improvements to this framework itself. They focus on making it more robust and scalable for real-world use.

Kai: One big improvement they point out is the scalability; because the gradient cost doesn't depend on the parameter count, you can drive unneeded elements toward zero in a thirty element alloy, which helps select both the right elements and their proportions <ref:2606.25502#pg3>.

Lev: From a hardware perspective, that means we don't waste time calculating things involving dozens of elements that are clearly not contributing to the desired property. The per iteration cost increased by only a factor of three when moving from three to thirty compositional parameters, which is quite good.

Kai: They also mention being able to search for dopants for existing functional materials by tracing continuous substitution pathways from a known compound, using the gradient progress to find the best additions.

Mira: That's a very practical improvement because it allows us to design modifications on top of a known material rather than starting completely from scratch.

Lev: But they also admit some limitations right away; they note that thermodynamic stability and lattice relaxation haven't been addressed yet, so those things need to be tackled in future work.

Kai: They even mention that the differentiability extends beyond just atomic positions to external parameters like applied magnetic and electric fields, which opens up a whole new avenue for design.

Conclusion: Mira: So to wrap up, this paper on "A Differentiable DFT-Based Framework for Inverse Materials Design" shows a way to use reverse mode automatic differentiation within KKR-CPA to perform inverse design by treating composition as continuous variables.

Kai: It gives us a first principles route from a target property to the material that realizes it, and it’s incredibly flexible because any computable quantity can be used as the objective function.

Lev: From my side, I see the biggest win in terms of feasibility for real hardware because of how they handle scaling compositional spaces—the gradient cost being independent of the number of elements is a big factor for iterative search.

Mira: It moves us past just using high-throughput screening or simple generative models and provides a mechanism where we can systematically explore composition space toward a goal, backed by first principles calculations.

Kai: It’s a powerful tool for materials science because it turns the inverse problem into an optimization problem that we can solve directly with gradient descent, which is much more direct than relying on training data.

Lev: One final thought is that while they’ve shown how to find compositions, the limitations they state are still there—specifically, they haven't yet addressed thermodynamic stability or lattice relaxation, so that’s where the next set of work needs to go.

Mira: That sounds like a solid conclusion for now. We’ll keep an eye on this framework as it evolves and see how we can apply it further in other areas of condensed matter physics.

Kohei Ishii, Hisazumi Akai, Tetsuya Fukushima, Hikari Shinya, Koji Inui

Institute for Solid State Physics, University of Tokyo · Department of Precision Engineering, Graduate School of Engineering, The University of Osaka Department of Materials DX Research Center, National Institute of Advanced Industrial Science and Technology Tsukuba · Calmarion Inc.

cond-mat.mtrl-sci, physics.comp-ph

Submitted: 2026-06-24

Updated: 2026-11-09

Comments: 26 pages, 6 figures, 2 tables; revised calculations and added experimental validation

Code: https://github.com/jax-ml/jax

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 91/100

The gist: Discovering solid-state materials with target properties remains a central challenge in computational materials science<ref:2606.25502#pg2> Here we develop a first-principles inverse-design…

Key concepts

KKR-CPA
This is a first-principles method used to study materials where atomic compositions are treated as continuous variables. It avoids the need for supercells and analytically averages substitutional disorder, making it suitable for optimization.
Reverse-mode Automatic Differentiation (AD)
This technique calculates the gradient of an objective function with respect to all input variables simultaneously in a single backward pass. Crucially, its computational cost does not depend on how many variables are being optimized, enabling efficient search in large compositional spaces.
Gradient-based Inverse Design
Instead of brute-force screening or relying on training data, this method uses the calculated gradients to iteratively move toward a target property. It guides the search directly from a desired outcome back to the necessary material composition.
Continuous Composition Variables
The framework treats atomic compositions not as discrete choices but as continuous values. This allows optimization algorithms like gradient descent to explore smooth pathways through compositional spaces, enabling the discovery of materials spanning dozens of elements.

Terminology

Summary

Discovering solid-state materials with target properties remains a central challenge in computational materials science<ref:2606. p. 025502#pg2> Here we develop a first-principles inverse-design framework, integrating reverse-mode automatic differentiation into KKR-CPA—the Korringa–Kohn–Rostoker method with the coherent potential approximation—where atomic compositions are continuous variables to be optimized<ref:2606. p. 025502#pg2> Reverse-mode AD yields gradients of objective functions with respect to composition at a cost independent of the number of candidate elements, enabling gradient-based optimization to identify materials from compositional spaces spanning dozens of elements<ref:2606. p. 025502#pg2> In this framework, any computable quantity can serve as the objective<ref:2606. p. 025502#pg2> We demonstrate this generality through two contrasting applications, magnetic alloys and halfmetals, yielding candidates such as (Lu0. p. 0553Yb0.447)(Co0.759Fe0.241)2Fe3 and FeZr(Sb0. p. 094Te0.06)<ref:2606. p. 025502#pg2> Our framework offers a physically grounded route from a target property to the material that realizes it<ref:2606. p. 025502#pg2>

Framework Foundation

The core of the method is integrating reverse-mode AD into KKR-CPA, treating atomic compositions as continuous variables for optimization<ref:2606.25502#pg2> The KKR-CPA method is a first-principles framework in which compositions are intrinsically continuous and differentiable<ref:2606. p. 225502#pg2> This approach requires no supercell, and its cost is nearly independent of the number of elements mixed on each site<ref:2606. p. 225502#pg2> The CPA performs the configurational average over substitutional disorder analytically, assigning each site continuous concentrations x(s)i specifying the probability that element i occupies site s<ref:2606. p. 225502#pg2> The total magnetic moment is defined as the integrated spin density µtot(x) = Z[ρ↑(r; x) − ρ↓(r; x)] dr<ref:2606. p. 225502#pg2>

Optimization Mechanism

Gradient-based inverse design follows the gradient of an objective straight toward a target, avoiding both the brute-force cost of screening and the training-data dependence of data-driven methods<ref:2606.25502#pg2> Automatic differentiation computes exact gradients with respect to many variables for essentially any differentiable objective, and now underpins gradient-based inverse design across fields<ref:2606. p. 125502#pg2> Reverse mode AD computes the gradient of a scalar output with respect to all input parameters in a single backward pass, at a cost independent of the parameter count<ref:2606. p. 125502#pg2> The workflow implements this as an iterative loop: KKR-CPA forward evaluation produces O(x), reverse-mode AD provides ∇xO, and the composition is updated<ref:2606. p. 225502#pg2>

Generality and Application

The framework proves its generality by optimizing contrasting targets, such as the total magnetic moment, µtot<ref:2606.25502#pg2> It also optimizes a target based on the shape of the density of states (DOS) to realize a half-metal<ref:2606. p. 225502#pg2> For example, in designing a half-Heusler XY Z compound, the optimization converges to FeZr(Sb0. p. 394Te0.06), which is close to the known half-metal FeZrSb<ref:2606. p. 325502#pg2> The framework can optimize for any target that KKR-CPA can compute, including electronic structure of disordered systems, magnetism and Curie temperatures<ref:2606. p. 225502#pg2>

Scalability and Robustness

The method scales to inverse design in large compositional spaces because the gradient cost is independent of the parameter count<ref:2606.25502#pg2> For a 30-element alloy, gradient-based optimization drives unneeded elements toward zero, selecting both the elements and their proportions<ref:2606. p. 225502#pg2> The per-iteration cost increased by only a factor of 3 when moving from 3 to 30 compositional parameters<ref:2606. p. 425502#pg2> Furthermore, the framework can search for dopants that most improve an existing functional material by tracing continuous substitution pathways from a known compound<ref:2606. p. 525502#pg2>

Future Directions

The present implementation uses the atomic-sphere approximation, but future work can incorporate a full-potential treatment to improve accuracy<ref:2606.25502#pg2> Thermodynamic stability and lattice relaxation are not yet addressed, but these limitations can be lifted in future work by differentiating the total energy with respect to atomic positions<ref:2606.25502#pg2> The same differentiability extends beyond atomic positions to external parameters such as applied magnetic and electric fields, pointing toward a framework that inverse-designs composition, lattice, and external fields together<ref:2606. p. 1

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. Gradient-based inverse design of materials can be performed directly from first principles calculations by integrating reverse-mode automatic differentiation into KKR-CPA to yield gradients of objective functions with respect to composition at a cost independent of the number of candidate elements. This enables gradient-based optimization to identify materials from compositional spaces spanning dozens of elements.

  2. AI systems can perform inverse design by treating composition as an intrinsically continuous variable, allowing them to optimize compositions by gradient descent and achieve this at no added computational cost relative to the number of parameters. This capability allows for the search for entirely unknown materials with the desired properties.

  3. The framework can be applied to design dopants for existing functional materials by tracing a continuous path from SmCo5 to the metastable PrFe5 using gradient progress, providing concrete doping levels backed by first-principles calculations.

Sources

Related papers