프리트레인된 머신러닝 인터액티브 포텐셜의 미세조정으로 구조 최적화 정확도 30% 향상
📝 Abstract
Accurate structural relaxation is critical for advanced materials design. Traditional approaches built on physics-derived first-principles calculations are computationally expensive, motivating the creation of machine-learning interatomic potentials (MLIPs), which strive to faithfully reproduce first-principles computed forces. We propose a fine-tuning method to be used on a pretrained MLIP in which we create a fullydifferentiable end-to-end simulation loop that optimizes the predicted final structures directly. Trajectories are unrolled and gradients are tracked through the entire relaxation. We show that this method consistently improves performance across all evaluated pretrained models; resulting in an average of roughly 32% reduction in prediction error. Interestingly, we show the process is robust to substantial variation in the relaxation setup, achieving negligibly different results across varied hyperparameter and procedural modifications.
💡 Analysis
Accurate structural relaxation is critical for advanced materials design. Traditional approaches built on physics-derived first-principles calculations are computationally expensive, motivating the creation of machine-learning interatomic potentials (MLIPs), which strive to faithfully reproduce first-principles computed forces. We propose a fine-tuning method to be used on a pretrained MLIP in which we create a fullydifferentiable end-to-end simulation loop that optimizes the predicted final structures directly. Trajectories are unrolled and gradients are tracked through the entire relaxation. We show that this method consistently improves performance across all evaluated pretrained models; resulting in an average of roughly 32% reduction in prediction error. Interestingly, we show the process is robust to substantial variation in the relaxation setup, achieving negligibly different results across varied hyperparameter and procedural modifications.
📄 Content
A central task in computational materials science is the identification of physically realizable atomic structures. In practice, this amounts to finding atomic configurations that correspond to local minima of the potential energy surface (PES), which maps atomic coordinates-given fixed species and electronic state-to potential energy. This work addresses how machine learning interatomic potentials (MLIPs) can be trained to more effectively predict relaxed states, without requiring additional expensive first-principles data.
Structural Relaxation and Its Cost. Stable structures are typically obtained through relaxation trajectories: iterative, gradient-based procedures that update an initial configuration toward lower-energy states using PES gradients. These gradients correspond to interatomic forces and are conventionally computed using density functional theory (DFT). While DFT provides accurate forces at the quantum-mechanical level, each evaluation is computationally expensive, with cost scaling steeply with system size, basis choice, and functional. As a result, full structural relaxations can require hours to days on high-performance computing systems, making relaxation a major bottleneck in high-throughput atomistic workflows.
MLIPs for Efficient Relaxation. To reduce this cost, machine learning interatomic potentials (MLIPs), also referred to as machine learning force fields, are trained to approximate DFT forces and energies (Batatia et al., 2022;Chen & Ong, 2022;Choudhary & DeCost, 2021;Deng et al., 2023;Yang et al., 2024a;Cheon et al., 2020;Musaelian et al., 2023;Dramko et al., 2025). Rather than directly predicting relaxed structures, MLIPs are typically used within iterative optimization loops to emulate DFT-driven relaxation at a fraction of the computational cost. In practice, they often serve either as full replacements for DFT during relaxation or as pre-relaxation tools that move structures close to equilibrium before final DFT refinement (Rossignol et al., 2023;Dramko et al., 2025). Although direct structure prediction models have received growing attention, MLIPs remain the dominant and most reliable approach for nontrivial structural domains (see Section 2.2).
Data Limitations and Motivation. A fundamental challenge in MLIP development is data scarcity. Each new training example requires costly first-principles calculations, resulting in datasets that are orders of magnitude smaller and less diverse than those in many other machine learning domains (Huang et al., 2023;Hörmann et al., 2025). Consequently, improving performance by simply scaling datasets is often impractical. This work instead focuses on extracting more value from existing data by altering how MLIPs are trained, rather than what data they are trained on. Contributions. Our primary contributions are as follows:
We introduce a full-trajectory, BPTT-based fine-tuning framework for MLIPs.
We provide ablation studies and analysis of the components involved in PES-level optimization.
We connect BPTT-based training to the theory of iterative maps and proxy functions.
We validate the approach across multiple structural domains and MLIP architectures.
MLIPs are trained to take as inputs structure snapshots1 and produce predictions for forces and formation energy.
A frequent direction of research in MLIPs has been graph neural networks (GNNs), which have led to many of the leading architectures (Batatia et al., 2022;Chen & Ong, 2022;Choudhary & DeCost, 2021;Deng et al., 2023;Yang et al., 2024a;Cheon et al., 2020;Musaelian et al., 2023). Another architecture of particular relevance to this paper is ADAPT (Dramko et al., 2025), which uses a points-in-space representation of atoms rather than a graphical one, and uses a Transformer encoder to predict atomic forces. A similar architecture appears in (Kreiman et al., 2025) and (Elhag et al., 2025).
Another common approach is the use of multi-layer perceptrons (and related base architectures). Appendix A.2 covers many of the variations used in literature.
Some works focused on improving MLIP performance have looked at using targeted or active learning to automatically identify areas of MLIPs that need to be trained (Sivaraman et al., 2020;Jacobs et al., 2025;Butler et al., 2024).
There has been growing interest in predicting relaxed atomic structures directly, without explicitly optimizing along the PES. However, because structural relaxation is a highly non-convex optimization problem, the data requirements for one-shot approaches are typically prohibitive (Sercu et al., 2021;Shu et al., 2018;Yang et al., 2024c). While limited success has been demonstrated in narrowly constrained settings, such as two-dimensional point defects (Yang et al., 2025), direct prediction remains challenging for nontrivial systems and consistently underperforms force-driven iterative ML relaxations on standard benchmarks (Kolluru et al., 2022;Yang et al., 2024c). As a result, iterative
This content is AI-processed based on ArXiv data.