Differentiable DFT: Inverse Materials Design via Gradient Composition Optimization ✛ ✎ ✦
Context
Searching for new functional materials is a central task in solid-state physics. First-principles methods, pioneered by quantum solid-state theory founders such as John Bardeen, can predict properties from composition, but the inverse problem—from property to composition—remained unsolved due to the enormous number of possible element combinations. Even with a fixed crystal lattice, the composition space is astronomically large. High-throughput screening requires huge resources, and approaches based on machine learning (surrogate models and generative design) are unreliable outside the training set. This work proposes a fundamentally different path—using gradient optimization within first-principles calculations, made possible by treating composition as continuous variables in the KKR-CPA method, where, unlike discrete supercells, the wave nature of electrons is described via Green's functions, and configurational averaging is done analytically.
Methods
The method is based on integrating numerical simulation using density functional theory with reverse-mode automatic differentiation. Instead of conventional plane-wave DFT, where swapping an atom changes the system discretely, the authors use the Korringa–Kohn–Rostoker method with the coherent potential approximation (KKR-CPA). This approach has its roots in the multiple scattering theory of Hans Bethe and uses Green's functions. Here, each crystallographic site can be occupied by a mixture of elements with continuous concentrations, and configurational averaging is done analytically. The computational graph is traversed twice: the forward pass computes the objective function (e.g., magnetic moment), and the backward pass yields exact gradients with respect to all concentration parameters, with the cost of this step independent of the number of parameters. The calculations employ Bloch's theorem (formulated by Felix Bloch) to describe electronic states in a periodic lattice and the Schrödinger equation with an exchange-correlation functional. The Pauli exclusion principle governs state occupation and spin polarization, crucial for magnetic materials.
Results
The method was tested in two contrasting scenarios. When maximizing the magnetization of the FeCoNi ternary alloy, gradient ascent stably converges to the well-known Slater–Pauling peak (Fe0.87Co0.13) from three different starting points. In a 30-dimensional space of elements from B to Te, the optimizer automatically suppresses all components except iron and cobalt, singling out the same optimal composition. For the multi-site LnT5 system with 43 parameters across three sublattices, a series of permanent magnet candidates were found, such as PrFe5 (12.15 μB/f.u.) and (Lu,Yb)(Co,Fe)2Fe3. In the second scenario, the target function was the density of states shape, and the optimization synthesized the half-metal FeZr(Sb0.94Te0.06)—close to the known FeZrSb—with a 97.6% suppression of states at the Fermi level in one spin channel. All candidates are metallic phases with pronounced magnetic properties. In all cases, each optimization step is backed by a full DFT calculation, and the final compositions were cross-checked for convergence.
Implications
This work demonstrates for the first time that combining KKR-CPA with automatic differentiation yields a universal first-principles inverse materials design tool. The fact that you can set any differentiable quantity as a target—from integral characteristics like magnetic moment to spectroscopic details of the electronic structure—opens the way to goal-oriented searches for materials with desired transport, optical, or catalytic properties. Moreover, the method requires no training data and avoids the extrapolation errors characteristic of machine learning models, which is crucial when exploring uncharted regions of phase space.
Future development
In the future, this approach can be extended to optimize not only composition but also lattice geometry and external parameters such as magnetic and electric fields. This would enable a unified computational platform for materials design “from target to substance,” where both composition and structure are jointly tailored. Developing full-potential calculation versions and incorporating thermodynamic stability are the next steps that will bring the method closer to real experimental challenges.
Impact
The method will impact the development of magnetic materials, especially high-performance permanent magnets and spintronic devices, as well as the search for new superconductors and thermoelectrics. Integration with calculations of electromagnetic properties, such as magnetic anisotropy and transport phenomena, will make materials design goal-oriented.
Next steps
Immediate tasks include implementing Curie temperature calculations via the Liechtenstein formalism, incorporating elemental cost targets into the objective function, and extending the method to strongly correlated systems.
Key open problems
This work directly addresses the fundamental problem of predictive materials design—minimizing blind screening. Using continuous composition variables combined with exact quantum field methods (Green's functions) bypasses the limitation of discrete supercells and enables exploration of spaces with dozens of parameters, previously inaccessible to first-principles methods.
The knowledge machine's view
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
The paper presents an inverse design method for materials that combines KKR-CPA with automatic differentiation for gradient-based composition optimization. The authors demonstrate it on maximizing magnetization in FeCoNi and searching for permanent magnets in the LnT5 system, recovering both known and new compositions.
Where its strength is
A strong point is the exact gradients with respect to continuous concentration parameters within first-principles DFT, which allows efficient exploration of high-dimensional composition spaces without surrogate models.
Where it could go next
- One could consider accelerating the convergence of gradient optimization with machine learning methods, similar to warm-starting for quantum calculations of molecules.based on: Warm Start for a Quantum Engine: How Machine Learning Accelerates Molecular Calculations arXiv:2607.09304
- It would be interesting to apply the approach to strongly correlated systems, where the KKR-CPA formalism could be extended beyond single-scattering, as in work on collective nuclear excitations.based on: The Atomic Nucleus: Not a Soloist, but a Choir of Voices arXiv:2605.01308
- The method might be extended to multicomponent alloys with multiple sublattices to search for materials with targeted magnetic anisotropy, using ideas from symmetry analysis of hidden magnetism.based on: How to Spot Hidden Magnetism by Symmetry arXiv:2605.27151
Worth trying
- Train a neural network surrogate on gradient trajectories to predict optimal compositions without full calculations.
- Use active learning to select starting points for optimization based on model uncertainty.
- Test the transferability of gradients across different crystal lattices to speed up screening.
Why this may matter
The method paves the way for rapid, targeted discovery of functional materials, cutting the time from problem definition to candidate synthesis, which could accelerate the development of new magnetic and electronic devices.
Nearby works in our archive
- Warm Start for a Quantum Engine: How Machine Learning Accelerates Molecular Calculationssimilarity 0.64
- The Atomic Nucleus: Not a Soloist, but a Choir of Voicessimilarity 0.63
- How to Spot Hidden Magnetism by Symmetrysimilarity 0.61
- How Clever Sampling Teaches Metal to Predict Strengthsimilarity 0.61
- How to Spot a Detail Without Overexposingsimilarity 0.60
- Quantum Dance in a Protective Cocoonsimilarity 0.60
The closest work in the archive sits at 0.64 by meaning — the usual level for our corpus.
Interesting fact: The Slater–Pauling peak, an empirical rule discovered in the 1930s, was independently “rediscovered” by the optimizer, which found the magnetization maximum in the Fe-Co system in just a few dozen steps, confirming the approach's validity.
Key numbers
- number of optimization parameters in the multi-site LnT5 system: 43
- maximum magnetic moment (PrFe5): 12.15 μB/f.u.
- spin polarization of the half-metal: 97.6%
- increase in computation time with a tenfold increase in parameters: only 3 times
- energy convergence accuracy: 10-6 Ry/atom