Source-linked AI summary
Physics Guided Deep Learning for Generative Design of Crystal Materials with Symmetry Constraints
Yong Zhao, Edirisuriya M. Dilanga Siriwardane, Zhenyao Wu, Nihang Fu, Mohammed Al-Fahdi, Ming Hu, Jianjun Hu
TL;DR
Crystal-material discovery covers a vast design space, while conventional methods are costly and existing generators face constraints in modeling physical rules and crystal symmetry. PGCGM combines symmetry-aware generation with physics-based losses; it reports higher optimization success than CubicGAN and strong match-rate improvements over both baselines, with large affine transformations posing a practical scaling constraint.
Problem
Crystal design spans a vast space, while physical and geometric complexity, limited data, and constraints on existing generators make effective generation difficult.
Source: §1 · pp. 1–3Method
PGCGM combines symmetry-aware crystal generation and self-augmentation with physics-based losses for atomic distances and structural symmetry.
Source: §1 · p. 3, Overall architecture of physics guided… · p. 13Results
93.5% (1,869) of generated materials were successfully optimized by DFT, versus 33.8% for CubicGAN; match rate exceeded FTCP by 4,200% and CubicGAN by 34.23%.
Source: §2 · pp. 5–6Takeaways & Limitations
The model's distance-based losses constrain atom spacing, while explicit cost functions encode the symmetry requirements of generated crystals.
Source: §2 · p. 9Converting base atom sites to full sites can produce large atom counts, particularly for space groups 227 and 225.
Source: §2 · p. 8
Abstract
from arXiv · showhide
Discovering new materials is a challenging task in materials science crucial to the progress of human society. Conventional approaches based on experiments and simulations are labor-intensive or costly with success heavily depending on experts' heuristic knowledge. Here, we propose a deep learning based Physics Guided Crystal Generative Model (PGCGM) for efficient crystal material design with high structural diversity and symmetry. Our model increases the generation validity by more than 700\% compared to FTCP, one of the latest structure generators and by more than 45\% compared to our previous CubicGAN model. Density Functional Theory (DFT) calculations are used to validate the generated structures with 1,869 materials out of 2,000 are successfully optimized and deposited into the Carolina Materials Database \url{www.carolinamatdb.org}, of which 39.6\% have negative formation energy and 5.3\% have energy-above-hull less than 0.25 eV/atom, indicating their thermodynamic stability and potential synthesizability.
1 Introduction
The introduction frames crystal design as a large, difficult search problem: experiments and conventional workflows are costly, while physical and geometric interactions complicate heuristic design. It presents PGCGM as a physics-guided generative model that incorporates crystal symmetry and physics-oriented training losses.
1 Introduction
- Only around 250,000 inorganic materials have been experimentally determined, a tiny portion of the vast design space, while trial-and-error discovery is time-consuming and labor-intensive.The introduction also stresses that complex atomic interactions make it difficult to enumerate design rules for heuristic approaches.Source: §1 · pp. 1–2
- High-throughput virtual screening cannot generate beyond known structural prototypes and is computationally intensive, while crystal-structure prediction methods are limited to relatively simple structures with few unit-cell atoms.Source: §1 · p. 2
- Generative models learn from known materials, but existing approaches have difficulty generating high-symmetry crystals and face limits in chemical-system, space-group, coordinate, and scale coverage.The introduction identifies four remaining challenges: learning physical constraints, precisely generating coordinates and lattice parameters, handling space-group distribution bias, and exploiting symmetry.Source: §1 · pp. 2–3
- PGCGM combines space-group affine transformation and self-augmentation with physics-oriented losses based on atomic pairwise distances and structural symmetry.Source: §1 · p. 3
- The model was evaluated against two baselines; DFT calculations validated its generated structures, with 93.5% of generated structures optimized successfully, a high success rate.Source: §1 · p. 3
2 Results and Discussion
PGCGM combines physics-based losses with symmetry constraints to generate diverse crystal structures, outperforming the reported baselines on generation and optimization evaluations. DFT calculations optimized 1,869 of 2,000 generated materials, and the study reports rediscovery, lattice-quality, and stable-material results across multiple space groups.
Dataset and model context
- PGCGM is evaluated on 42,072 ternary materials spanning 20 space groups, against FTCP and CubicGAN; its generator uses atom-distance and symmetry-related losses to improve structural quality.The authors state that the model could apply to other space groups given sufficient training samples.Source: §2 · pp. 3, 9
Generation performance
- PGCGM significantly outperforms FTCP and CubicGAN on property-distribution metrics, while PGCGM achieves the best diversity score despite generating more readable CIFs than FTCP.PGCGM+dist+coor is selected for later analysis because it has better property-distribution performance on average than PGCGM and PGCGM+dist.Source: Table 1 · p. 4, §2 · p. 4
Bayesian optimization
- PGCGM’s BOWSR match rate exceeds FTCP’s by 4,200% and CubicGAN’s by 34.23%, although FTCP’s very low match rate accompanies its analyzed structures all having space group P1.The study selected 2,000 materials per method for BOWSR analysis, with PGCGM’s sample evenly drawn across 20 space groups.Source: §2 · pp. 4–5, Table 2 · p. 4
Rediscovery
- At 60 million sampled materials, PGCGM rediscovered 52.0% of MIO and 43% of TST; the training-set rate reached 42.6% after 35 million samples.The authors attribute the lower TST rediscovery rate to different space-group proportions in the training and test datasets.Source: §2 · p. 5, Figure 1 · p. 5
DFT validation
- 1,869 of 2,000 generated crystals (93.5%) optimized successfully; 39.6% had negative formation energy and 106 of 1,579 had energy above hull below 0.25 eV/atom.Three materials had energy above hull of 0 eV/atom.Source: §2 · pp. 6–7, Figure 2 · p. 6
Lattice parameters
- PGCGM’s generated lattice lengths fit DFT-relaxed lengths better than CubicGAN by R2 across 1,869 materials and by both R2 and RMSE for cubic materials.The cubic comparison uses 293 PGCGM and 14,432 CubicGAN materials; PGCGM’s mean relative error is closer to zero, while CubicGAN’s outliers cluster near 150%.Source: §2 · p. 7, Figure 4 · p. 7
Structure post-processing
- Clustering and merging reduce crowded generated atom counts, but can change formulas even when the space group is retained.The text notes that conversion from base to full atom sites can produce large atom counts, especially for space groups with large affine matrices such as 227 and 225.Source: §2 · pp. 7–8, Table 3 · p. 8
Stable-material case studies
- DFT case studies identify Mg2GaIr, SrYO6, and ZnTe2S6 as thermodynamically stable; SrYO6 is a half-metal, and all three meet their reported mechanical and 0 K dynamical stability criteria.Mg2GaIr and ZnTe2S6 are metallic, while SrYO6 is ferromagnetic with a 1 µB moment and a 3.09 eV band gap for spin-up electrons.Source: Figure 3 · p. 6, §2 · p. 9
Methods
PGCGM represents ternary crystals with element identities, base atom sites, lattice parameters, and space-group symmetry. It combines symmetry-aware generation with physics-based losses and evaluates generated structures using structural metrics and DFT validation.
Problem statement and notations
- PGCGM models ternary materials as element identities, base atom sites, lattice parameters, and space-group symmetry, and generates base sites and lattice parameters from noise, elements, and space group.The training dataset is restricted to ternary materials, with one base atom site per element in the stated representation.Source: Problem statement and notations · p. 10
Overall architecture of physics guided crystal generative model
- Symmetry consistency and inter- and intra-atom distance losses guide the generator, while self-augmentation creates equivalent samples by transforming base sites under space-group symmetry.The symmetry losses enforce consistency between the full coordinates derived from different base-site sets; distance losses constrain atoms to a specified range.Source: Overall architecture of physics guided… · pp. 11–13
Implementation details
- High-symmetry structures can yield many unit-cell atoms, so post-processing clusters nearby atoms of the same element and merges atoms within clusters.The maximum allowed clustering distance is 1.2 times the atom-radius sum.Source: Overall architecture of physics guided… · p. 13
- Because exact coordinate-uniqueness checks are nondifferentiable and time-consuming, the implementation averages symmetry-transformed coordinates and uses the resulting full-coordinate representations in consistency constraints.Three sets of base sites are transformed using affine operations; their averaged full coordinates are expected to correspond to the same material.Source: Implementation details · pp. 14–15
Evaluation Metrics
- Evaluation measures validity, property-distribution distance, composition diversity, and DFT-based structural relaxation, thermodynamic stability, and synthesizability.The property-distribution metric uses Wasserstein distance for minimum atom distance, maximum atom distance, and density; the passage describes these structural metrics as indirect, weak criteria relative to DFT validation.Source: Evaluation Metrics · p. 15