A Computed Combinatorial Landscape for Magnetically Driven Polar Distortion

In some crystals, magnetic order physically moves the atoms. When spins align, the crystal can lower its energy by shifting ions so that favorable magnetic bonds get shorter and unfavorable ones get longer. This effect is called exchange striction. If those shifts break the crystal's mirror-through-a-point symmetry (inversion symmetry), the material becomes electrically polar because it is magnetic. The double perovskite Pb₂CoOsO₆ is a clean example: it is centrosymmetric at room temperature and becomes polar at 45 K to relieve magnetic frustration (Princep et al.). Moderate pressure suppresses the moments and restores the symmetric structure (Jiao et al.).

Materials like this have been found by luck or by hand. Each candidate requires small energy differences computed across several magnetic configurations, and that cost scales badly. Protein engineering faced the same combinatorial wall and solved part of it with a specific recipe. Wittmann et al. built an exhaustively measured landscape (GB1, all 160,000 four-site variants). They used it to show that greedy one-mutation-at-a-time search fails on epistatic landscapes, meaning landscapes where mutations interact non-additively. They also showed that a cheap physics filter, which removes sequences that will not fold before any model is trained, makes learned search efficient (Wittmann et al.).

This proposal builds the materials analogue of GB1. We will compute, exhaustively and from first principles, the size of the magnetically driven polar distortion across every variant of a four-site chemical substitution library built around Pb₂CoOsO₆. We then use that ground truth to test a transfer hypothesis: that a cheap, zero-shot estimate of the exchange-striction force, computed without relaxing the structure, plays the role Rosetta stability scores play for proteins.

Core Claim

We claim two things. First, magnetoelastic distortion over a substitution library is a cliff-edge, epistatic landscape. Most variants are "holes," in the sense used by Chen et al.. In those variants, magnetic order collapses, the antiferromagnetic state is not the ground state, or the lattice is already polar without magnetism. The useful variants depend on combinations of substitutions, so single-site walks get stuck. Second, the holes can be predicted cheaply. Take the undistorted, high-symmetry structure, freeze in a small displacement along the polar mode, and compute the change in exchange constants. The resulting linear force on the polar mode is the $\partial J/\partial r$ quantity that Wang et al. identify as the driver.

Using this force estimate as a filter that decides which variants get expensive training labels, as in focused-training machine-learning-directed evolution (ftMLDE), should find the top variants with a small fraction of the full computation. What is new is not "apply a protein optimizer to materials." What is new is a ground-truth landscape plus a test of which protein-engineering lessons actually transfer: landscape ruggedness, filter-focused training, and proximal search.

The Landscape

The scaffold is the A₂BB′O₆ double perovskite in a supercell with two inequivalent A sites arranged in layered order. This connects to the combinatory A-site ordering of Zhao et al., who built such compounds by hand rather than by search. The four substitutable sites, each with five options, are as follows. Each A site takes Pb, Ba, Sr, Ca or Cd. The B site takes Co, Ni, Mn, Fe or Cu. The B′ site takes Os, Re, Ir, Mo or W. This gives 625 variants. The library deliberately contains nonmagnetic B′ choices such as d⁰ W⁶⁺, so physically meaningful holes are built in, and it contains the known parent, Pb₂CoOsO₆.

Each variant is computed with the standard DFT+U approach used in this literature. Density functional theory (DFT) is the workhorse quantum-mechanical method for computing material energies, and the "+U" term corrects its treatment of localized electrons. For each variant we relax the structure in the lowest-energy collinear antiferromagnetic configuration and in a ferromagnetic reference, starting from the centrosymmetric structure with a small polar seed. The label is the magnetically driven part of the polar instability:

$$y = \Delta E_{\mathrm{pol}}^{\mathrm{AFM}} - \Delta E_{\mathrm{pol}}^{\mathrm{FM}},$$

Here $\Delta E_{\mathrm{pol}}^{X}$ is the energy gained by relaxing from the centrosymmetric to the polar structure while spins are held in configuration $X$. A negative $y$ means magnetism specifically favors polarity. Variants that are polar even in the ferromagnetic reference are flagged as ordinary ferroelectrics rather than magnetically driven ones. We also record the polar-mode amplitude and the magnetic moments.

The zero-shot proxy needs no relaxation. We evaluate the exchange constants $J_{ij}$ at the centrosymmetric structure and at one frozen polar displacement $u$. We then form

$$\hat{F} = -\sum_{i<j} \frac{\partial J_{ij}}{\partial u}\, \langle \mathbf{S}_i\cdot\mathbf{S}_j\rangle_{\mathrm{AFM}},$$

which is the first-order force on the polar mode from the Heisenberg model $\hat H=-\sum J_{ij}\mathbf{S}_i\cdot\mathbf{S}_j$. The proxy also checks that local moments survive.

Its cost relative to full relaxation is roughly a handful of static calculations, compared with many relaxation steps per spin configuration. The proxy deliberately ignores relaxation. The literature says ignoring relaxation can give the wrong answer: Bhat et al. show static calculations pick the wrong Fe₂P phase. Whether the proxy is still good enough as a filter, as opposed to a predictor, is the empirical question.

Before trusting the landscape, we run a sanity gate. The parent compound must come out polar in its antiferromagnetic state and centrosymmetric in the ferromagnetic reference. Compressing its lattice must reduce the moments and remove the distortion, which qualitatively reproduces Jiao et al. If DFT+U cannot reproduce the parent, we report that as a finding and switch the scaffold to an insulating exchange-striction system, because metallic 5d osmates are a hard case for these methods.

Research Program

The program has three parts. The first is building and releasing the 625-variant landscape with all intermediate quantities: exchange constants, moments, relaxed structures and the proxy.

The second is characterizing the landscape the way GB1 was characterized. We will decompose the fitness variance into additive, pairwise and higher-order site interactions, measure the fraction of holes, and count local optima under single-site moves. This tells us whether the protein analogy holds at all. If the landscape turns out nearly additive, that is a publishable negative and changes the recommended design strategy.

The third is replaying optimization campaigns on the landscape under fixed label budgets. Every method draws labels from the precomputed truth. We then scale to a seven-site library of about 78,000 variants, built by adding substitutable sites in a larger supercell. There exhaustive DFT is infeasible and the best method from the replay is run prospectively.

On choosing methods to transfer, we are explicit about the constraints. Energy Rank Alignment (ERA) aligns a pretrained protein language model (ESM3) with fitness rankings (Ibarraran et al.). Plug & Play Directed Evolution (PPDE) multiplies a pretrained evolutionary prior by a supervised model (Emami et al.). Both depend on a pretrained prior over plausible sequences, and no such prior exists for magnetic substitution chemistry. We therefore do not port them wholesale. We borrow only what makes sense without a prior: the pairwise ranking loss from ERA, and the distance-regularized search of Proximal Exploration (PEX) (Ren et al.), anchored at the experimentally realized parent.

This also bears on a tension in the file. Ibarraran et al. found that structural conditioning did not consistently help directed evolution, while Wittmann et al. found a structural filter decisive. Our filter ablation tests which result holds for magnetoelastic landscapes.

Baselines

The naive reference is uniform random sampling of variants under the same label budget. It answers whether any method beats picking compounds blindly.

The traditional chemist's strategy is a greedy single-site walk. Start at the parent, compute all single substitutions, move to the best, and repeat until the budget runs out. On GB1 this approach reached the global optimum only 1.2% of the time (Wittmann et al.). Whether it fails as badly here is itself a result.

The proxy-only ranker sorts all variants by $\hat F$ and labels the top ones with no learning. It is the cheapest physics-based competitor and the most dangerous one for our claim.

The strongest existing method is ftMLDE as implemented in the MLDE package (github.com/fhalab/MLDE). It encodes variants with one-hot and physicochemical descriptors and trains regressors (1D convolutional networks and boosted trees) on a training set focused by a zero-shot score. We substitute $\hat F$ for Rosetta stability scores. The reported reference is that ftMLDE reached the GB1 global maximum 92% of the time with a total screening budget of 470 variants, and was 8 times more likely than greedy evolution to find it at 80 variants (Wittmann et al.). There is no reported result on Lacuna for any optimizer on a magnetoelastic or exchange-striction landscape. Beating ftMLDE therefore means beating our own faithful reimplementation of it on our landscape, not beating the GB1 number.

PEX is run anchored at Pb₂CoOsO₆ with its distance penalty, using a convolutional surrogate because the mutation-factorization network PEX also proposes is defined for long sequences. Its reported overall normalized score was 0.98 against 0.43 for AdaLead on the TAPE oracle suite (Ren et al.). Again, there is no reported result on Lacuna for our domain.

The ablation that isolates the key ingredient is ftMLDE with the training set drawn at random instead of focused by $\hat F$. Everything else is held fixed: encoding, models, budget, and hyperparameter search done by nested cross-validation inside the labeled set. A second ablation swaps the regression loss for ERA's pairwise ranking loss.

All methods get identical budgets of 24, 48, 96 and 192 labels. The $\hat F$ evaluations are charged to the filtered methods at their measured cost in core-hours relative to a full label, and we run 100 random seeds. Our method counts as winning if filter-focused learning finds a variant in the true top 1% more often than both ftMLDE-without-filter and the proxy-only ranker at equal total compute, with non-overlapping bootstrap intervals.

Why not just…?

Why not just substitute one element at a time, as chemists do? On epistatic protein landscapes this strategy almost never reaches the optimum (1.2% on GB1). Frustration-driven polarity in Pb₂CoOsO₆ depends on two sublattices balancing each other, which suggests interacting sites. But that is an argument, not evidence. The epistasis decomposition and the greedy baseline settle it directly. If greedy matches the learned methods, we will say the landscape is smooth and the machine-learning machinery is unnecessary.

Why not just compute everything with DFT? For 625 variants we do exactly that, because a ground truth is required. The problem is scaling. A single spin-resolved mechanical study of a 72-atom metal-organic framework took about 25 CPU-years (Vanpoucke et al.), and adding three sites multiplies the library roughly 125-fold. The seven-site prospective campaign exists to show the regime where exhaustive computation stops being an option.

Why not just rank by the cheap exchange-striction force and skip learning? This is the real threat. If $\hat F$ alone ranks the true top variants well, the contribution reduces to "use the force estimate," which is still useful but simpler. The literature cuts both ways. Fe₂P shows that unrelaxed calculations can mislead about structural ground states, while the $\partial J/\partial r$ picture says the force term is exactly what drives the distortion. The proxy-only baseline and the Spearman correlation between $\hat F$ and $y$ decide this. We commit to reporting it even if it beats the learned methods.

Why not just use a generative crystal model such as MOFDiff? MOFDiff assembles frameworks from modular building blocks and cut the number of expensive simulations needed for carbon-capture targets by 75%. However, it fixes lattice parameters and ignores spins, so it cannot represent a distortion that magnetic order creates. The machine-learning spin-crossover classifiers (equivariant screening; transition temperatures) model only local, single-molecule spin switching and do no active search. Neither approach targets collective, symmetry-breaking striction, which is why a fixed-scaffold library is the right first object.

Decisive Checks

The work stands or falls on four measurements. The first is the parent gate: the computed landscape must reproduce polarity in the antiferromagnetic state, symmetry in the ferromagnetic reference, and loss of polarity under compression for Pb₂CoOsO₆.

The second is landscape statistics: the share of variance from interaction terms, the hole fraction, and the number of single-move local optima. Substantial interaction variance plus multiple local optima confirms the analogy. A near-additive landscape refutes it.

The third is filter value. We compare top-1% hit rates and the best value of $y$ found against label budget for focused versus unfocused ftMLDE and the proxy-only ranker, at equal compute. If focusing does not help, the protein lesson does not transfer. That would side with Ibarraran et al.'s finding that structure-based conditioning is not consistently helpful.

The fourth is prospective validation on the seven-site library. The top proposed variants are fully recomputed, and we report how many show a negative $y$, meaning magnetism specifically favors polarity, larger in magnitude than the parent's. A positive outcome yields candidate polar magnets for synthesis. Either outcome yields the first public benchmark for search over magnetically driven distortion.

References

  1. Magnetically driven loss of centrosymmetry in metallic <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mrow><mml:msub><mml:mi>Pb</mml:mi><mml:mn>2</mml:mn></mml:msub><mml:msub><mml:mi>CoOsO</mml:mi><mml:mn>6</mml:mn></mml:msub></mml:mrow></mml:math>
  2. Coupled magnetic and structural phase transitions in the antiferromagnetic polar metal <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mrow><mml:msub><mml:mi>Pb</mml:mi><mml:mn>2</mml:mn></mml:msub><mml:mi>Co</mml:mi><mml:mi>Os</mml:mi><mml:msub><mml:mi mathvariant="normal">O</mml:mi><mml:mn>6</mml:mn></mml:msub></mml:mrow></mml:math> under pressure
  3. Machine Learning-Assisted Directed Evolution Navigates a Combinatorial Epistatic Fitness Landscape with Minimal Screening Burden
  4. Generalists vs. Specialists: Evaluating LLMs on Highly-Constrained Biophysical Sequence Optimization Tasks
  5. First-principles study of the lattice and electronic structure of<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:mi mathvariant="normal">Tb</mml:mi><mml:msub><mml:mi mathvariant="normal">Mn</mml:mi><mml:mn>2</mml:mn></mml:msub><mml:msub><mml:mi mathvariant="normal">O</mml:mi><mml:mn>5</mml:mn></mml:msub></mml:mrow></mml:math>
  6. A combinatory ferroelectric compound bridging simple ABO3 and A-site-ordered quadruple perovskite
  7. Role of zero-point effects in stabilizing the ground state structure of bulk Fe2P
  8. Efficient, Few-shot Directed Evolution with Energy Rank Alignment
  9. Plug & play directed evolution of proteins with gradient-based discrete MCMC
  10. Proximal Exploration for Model-guided Protein Sequence Design
  11. Quasi-1D physics in metal-organic frameworks: MIL-47(V) from first principles
  12. MOFDiff: Coarse-grained Diffusion for Metal-Organic Framework Design
  13. Discovery of Spin-Crossover Materials with Equivariant Graph Neural Networks and Relevance-Based Classification
  14. Machine Learning Prediction of the Experimental Transition Temperature of Fe(II) Spin-Crossover Complexes