LightPFP: A Lightweight Route to Ab Initio Accuracy at Scale

TL;DR

LightPFP presents a data-efficient distillation framework that converts a high-fidelity universal MLIP into a lightweight, task-specific MLIP, enabling rapid production-scale simulations without exhaustive DFT labeling. By using a pre-trained, data-rich universal teacher (PFP) to generate labeled data and a compact pre-trained MTP student that is fine-tuned on teacher data, LightPFP achieves 1–2 orders of magnitude faster inference and up to 3 orders of magnitude faster model development compared to conventional DFT-based ts-MLIPs. The framework demonstrates competitive accuracy across solid-state electrolytes, high-entropy alloys, and reactive surface processes, with targeted transfer learning (as few as 10 high-accuracy DFT points) correcting systematic teacher errors in MgO melting-point predictions. The approach blends data-efficient distillation, active learning for complex chemistry, and few-shot transfer learning to deliver scalable, high-fidelity MLIPs suitable for large-scale molecular dynamics and materials discovery. This methodology promises broad applicability to diverse materials problems where rapid, precise, and scalable atomistic modeling is essential.

Abstract

Atomistic simulation methods have evolved through successive computational levels, each building upon more fundamental approaches: from quantum mechanics to density functional theory (DFT), and subsequently, to machine learning interatomic potentials (MLIPs). While universal MLIPs (u-MLIPs) offer broad transferability, their computational overhead limits large-scale applications. Task-specific MLIPs (ts-MLIPs) achieve superior efficiency but require prohibitively expensive DFT data generation for each material system. In this paper, we propose LightPFP, a data-efficient knowledge distillation framework. Instead of using costly DFT calculations, LightPFP generates a distilled ts-MLIP by leveraging u-MLIP to generate high-quality training data tailored for specific materials and utilizing a pre-trained light-weight MLIP to further enhance data efficiency. Across a broad spectrum of materials, including solid-state electrolytes, high-entropy alloys, and reactive ionic systems, LightPFP delivers three orders of magnitude faster model development than conventional DFT-based methods, while maintaining accuracy on par with first-principles predictions. Moreover, the distilled ts-MLIPs further sustain the computational efficiency essential for large-scale molecular dynamics, achieving 1-2 orders of magnitude faster inference than u-MLIPs. The framework further enables efficient precision transfer learning, where systematic errors from the u-MLIP can be corrected using as few as 10 high-accuracy DFT data points, as demonstrated for MgO melting point prediction. This u-MLIP-driven distillation approach enables rapid development of high-fidelity, efficient MLIPs for materials science applications.

Figures (10)

• Figure 1: A standard "food chain" of atomistic calculation methods.
• Figure 2: Schematic diagram of LightPFP.
• Figure 3: Comparison of data efficiency between fine-tuned pretrained and scratch-trained student models.
• Figure 4: (a) Molecular dynamics (MD) computational speed with Li6PS5Cl as a function of number of atoms for three MLIPs: PFP, LightPFP (MTP), and MACE. (b) Trade-off between the overall time spent on MLIP building for Li6PS5Cl, including data collection and model training, and MD computational speed for PFP, LightPFP, MACE, and MTP-DFT. Inset: the total time cost to complete both MLIP building and a 10 ns MD simulation of a 10,000-atom system With PFP, LightPFP, MACE, and MTP-DFT.
• Figure 5: Parity plot comparing atomic forces predicted by MLIPs to DFT reference values (a) PFP; (b) LightPFP; (c) MACE and (d) MTP-DFT