Constrained Diffusion for Accelerated Structure Relaxation of Inorganic Solids with Point Defects
Proposes a primal-dual constrained diffusion model for realistic inorganic defect structures, outperforming existing methods with lower RMSD and forces.
Key Findings
Methodology
This paper introduces a constrained diffusion framework based on the primal-dual algorithm integrated into the score-based diffusion model. It incorporates multiple physical constraints—geometric, distributional, and force-based—via Lagrangian multipliers and neural surrogate models. The approach employs a multi-stage projection strategy to ensure physical plausibility while maintaining diffusion dynamics. The neural force predictor guides the soft force constraint, and the optimization iteratively refines the generated structures. Experiments on Bi2Te3 demonstrate superior performance in generating stable, physically grounded defect configurations, with significant improvements in RMSD, RDF similarity, and force metrics compared to baselines.
Key Results
- Across six defect configurations in Bi2Te3, the proposed model achieved an average RMSD of 0.90 Å, outperforming conditional diffusion (1.86 Å) and projected diffusion (2.54 Å). RDF similarity was doubled relative to baseline models, indicating better geometric fidelity. Total forces were reduced by several orders of magnitude, reaching 10^-1 eV/Å, signifying structures close to energy minima. These results confirm the model's ability to generate physically realistic defect structures with high diversity and stability.
- Compared to unconstrained models, the method maintains diffusion-driven diversity while enforcing physical constraints, avoiding gradient misalignment issues common in projected diffusion. The final structures exhibit lower RMSD and RDF deviations, demonstrating effective constraint handling without sacrificing generative quality. Ablation studies highlight the importance of the primal-dual optimization and neural force surrogate in achieving these results.
- The model generalizes well across different defect types, including vacancies and anti-sites, showing robustness and adaptability. Its efficiency and accuracy suggest promising applications in high-throughput defect engineering, accelerating material discovery and optimization processes in thermoelectric and semiconductor materials.
Significance
This work addresses a fundamental bottleneck in computational materials science—efficiently generating physically valid defect structures without expensive DFT calculations. By integrating physics-based constraints into a deep generative diffusion process, the approach offers a scalable, reliable alternative for high-throughput screening. Its success in Bi2Te3 demonstrates potential for broader application across diverse inorganic materials, enabling rapid exploration of defect-property relationships. The framework bridges the gap between data-driven generative models and physical realism, paving the way for smarter materials design and accelerated innovation in thermoelectrics, semiconductors, and beyond. The methodology also opens avenues for integrating complex multi-physics constraints into deep generative models, fostering more accurate and physically consistent simulations.
Technical Contribution
The key technical innovation lies in embedding primal-dual optimization within the score-based diffusion framework, enabling the enforcement of complex, non-convex physical constraints during generation. The use of neural surrogate models for force prediction allows soft constraint incorporation without prohibitive computational costs. The multi-stage projection approach, based on augmented Lagrangian duality, effectively relaxes the assumption of intermediate feasibility, focusing constraint enforcement on the final structure. This method overcomes limitations of prior constrained diffusion models that rely on noisy constraint evaluations, which often lead to gradient misalignments. The framework provides theoretical guarantees of convergence to feasible solutions while maintaining the diversity and realism of generated structures, representing a significant advance in physics-aware generative modeling.
Novelty
This is the first work to integrate primal-dual constrained optimization directly into score-based diffusion models for atomic structure generation. Unlike previous approaches that either impose constraints post hoc or rely on conditioning, this method dynamically enforces multiple complex physical constraints during the diffusion process. The combination of neural force surrogates, multi-stage Lagrangian optimization, and a relaxed terminal constraint formulation constitutes a novel paradigm, enabling the generation of physically plausible defect structures at scale. It bridges the gap between deep generative modeling and rigorous physical constraints, setting a new standard for scientific structure synthesis.
Limitations
- The approach depends heavily on the accuracy of neural surrogate models for force prediction; inaccuracies here can lead to less physically realistic structures. The computational cost remains high due to iterative Lagrangian optimization, especially for larger systems or more complex constraints. The current framework is validated primarily on Bi2Te3; its generalization to other materials or defect types requires further investigation. Additionally, extreme defect configurations or highly non-equilibrium states may challenge the model's robustness. Future work should focus on improving surrogate accuracy, reducing computational overhead, and extending applicability to broader material classes.
Future Work
Future directions include integrating multi-scale physical models to handle larger and more complex systems, improving surrogate force models via active learning, and exploring end-to-end training strategies for better constraint satisfaction. Extending the framework to other classes of materials, such as oxides or complex alloys, will broaden its impact. Additionally, incorporating experimental feedback and real-world defect data can enhance model reliability. Developing more efficient optimization algorithms and parallelization techniques will reduce runtime costs, facilitating industrial adoption. Ultimately, this work aims to establish a comprehensive, physics-informed generative platform for accelerated materials discovery.
AI Executive Summary
点缺陷在无机材料性能中扮演着关键角色，尤其在热电材料如Bi2Te3中，缺陷的配置和浓度直接影响其电热性能。然而，传统的第一性原理模拟如DFT虽然精确，却因计算成本高昂，难以实现大规模高通量筛选。近年来，深度学习中的扩散模型因其强大的生成能力逐渐成为替代方案，但在严格满足物理约束方面仍存在挑战。本文提出一种结合 primal-dual 算法的受约束扩散模型，有效融合几何、分布和力学约束，确保生成的缺陷结构符合实际物理规律。
该方法在Score-based扩散框架基础上引入多重约束，通过Lagrangian乘子和神经代理模型实现多目标优化。其核心思想是利用多阶段投影策略，避免梯度偏差，确保最终生成结构的物理合理性。实验在Bi2Te3的六种缺陷配置中表现出优异性能，RMSD最低（0.90 Å），RDF一致性提升两倍，结构稳定性显著增强。这一突破性技术不仅提升了生成质量，也极大降低了模拟成本，为高通量材料设计提供了新工具。
未来，模型有望扩展到多尺度、多物理场环境，推动材料科学的智能化发展。尽管如此，模型在极端缺陷状态和复杂约束下的表现仍需优化，算法效率也需提升。总体而言，该研究为科学计算中的受约束生成提供了新的理论基础和实践路径，有望加速新材料的发现与应用。
Deep Dive
Abstract
Point defects affect material properties by altering electronic states and modifying local bonding environments. However, high-throughput first-principles simulations of point defects are costly due to large simulation cells and complex energy landscapes. To this end, we propose a generative framework for simulating point defects, overcoming the limits of costly first-principles simulators. By leveraging a primal-dual algorithm, we introduce a constraint-aware diffusion model which outperforms existing constrained diffusion approaches in this domain. Across six defect configuration settings for Bi2Te3, the proposed approach provides state-of-the-art performance generating physically grounded structures.