Worse Forces Can Lead to Better Structures: Trajectory-Level Fine-Tuning of Machine Learning Interatomic Potentials
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Condensed Matter. The day in condensed matter: materials, phases and the measurements under them.
Kai: We're looking at the paper: "Worse Forces Can Lead to Better Structures".
Mira: Here's the gist of the paper. Accurate structural relaxation is critical for advanced materials design because traditional physics-derived methods are computationally expensive, motivating machine-learning interatomic potentials (MLIPs) which strive to faithfully reproduce first-principles computed forces The authors propose a fine-tuning method to be used on a pretrained MLIP in which they create a fully differentiable end-to-end simulation loop that optimizes the predicted final structures directly This method consistently improves performance across all evaluated pretrained models, resulting in an average of roughly 32% reduction in prediction error The core methodology involves a trajectory-level training strategy The authors adopt a trajectory-level training strategy where they unroll full relaxation trajectories and apply backpropagation through time (BPTT) to update model parameters based on the quality of the final relaxed structure rather than intermediate force errors This reframes MLIP training as a problem of optimizing the outcome of the relaxation process itself, allowing one data modality to supervise another Specifically, they define a frame as a single DFT/MLIP force-prediction step followed by a corresponding structural-update step The rollout is defined as Xb F = frame ◦ · · · ◦ frame
z: × k (X0) for k steps The training signal is built through coupled optimization levels The work centers on two distinct optimization processes: the MLIP training and the PES-level optimization of atomic coordinates during structural relaxation The key idea is to explicitly couple these two levels during training, treating parameter updates of the MLIP as MLIP training and coordinate updates as PES-level optimization They investigate various PES-level step-size controls, including a single learned scalar step size and a small neural network that predicts the step size dynamically during the relaxation trajectory The gains come from learning structural bias rather than physical dynamics The analysis suggests that BPTT fine-tuning is not improving the physical fidelity of force predictions nor optimizing the underlying relaxation dynamics in a conventional sense Instead, it learns a higher-level structural bias of the dataset by constructing a proxy iterative map that preserves key properties of the true PES, such as fixed points and their associated basins of attraction This is achieved because the objective is not to improve adherence to PES-level descent dynamics but to learn a map that matches fixed points and relevant basins-of-attraction to the PES Validation across diverse structural domains confirms generalizability The approach is validated across three datasets: crystal defects, pure crystals, and catalysts, showing that BPTT fine-tuning improves performance in all cases Qualitative analysis shows that the defect centers exhibit the most pronounced improvements in the resulting defect structures The results support a pragmatic interpretation that BPTT can reliably improve the trajectory-level behavior of MLIPs across many applied regimes, even if it does not "solve the physics" The final recommendation is a staged training pipeline The authors suggest treating BPTT as the final stage of a staged training pipeline: start from a pretrained, broadly applicable MLIP; refine it with supervised structure-to-force training on the target domain; then apply trajectory-level BPTT to shape the long-horizon relaxation This framing is supported by experimental results showing that BPTT fine-tuning is not an effective standalone learning paradigm but a powerful fine-tuning tool The benefit is not confined to extremely narrow tasks, as the pure crystal dataset exhibits substantial diversity spanning orders of magnitude in system size and covering a broad range of chemistries While BPTT does not "solve the physics," it can reliably improve trajectory-level behavior across many applied regimes The trade-off in high-throughput or computationally heavy studies of specific structure-classes is favorable because improved relaxation fidelity reduces the need for expensive DFT evaluations The authors conclude that BPTT fine-tuning lowers the data requirements for producing an effective domain-specific MLIP, addressing a common bottleneck in practical deployment REFERENCE arXiv:2512.
Paper summary: Kai: So, to recap what we just talked about, this paper by Dramko et al. is proposing a fine-tuning method for pretrained machine learning interatomic potentials. They claim that instead of just training the force predictor, you can use a fully differentiable end-to-end simulation loop that optimizes the final structures directly.
Mira: Essentially, they show that this approach consistently improves performance across all the pretrained models they tested, achieving an average reduction in prediction error of about thirty-two percent <ref:2512.01067#pg1,consistently improves performance across all>.
Lev: The key methodology involves a trajectory level training strategy where they unroll full relaxation trajectories and apply backpropagation through time, or BPTT, to update model parameters based on the quality of the final relaxed structure.
Kai: They set up two distinct optimization processes: first, optimizing the machine learning interatomic potential parameters during training, and second, optimizing atomic coordinates along the potential energy surface during structural relaxation.
Mira: The key idea here is explicitly coupling these two levels during training to avoid ambiguity between MLIP training and PES level optimization.
Lev: They investigate various ways to control the PES level step size, looking at a single learned scalar step size or a small neural network that predicts the step size dynamically throughout the relaxation trajectory.
Kai: The analysis suggests that this BPTT fine-tuning isn't just about improving physical fidelity of force predictions, but learning a structural bias of the dataset by constructing an iterative map.
Mira: This map preserves key properties of the true potential energy surface, like fixed points and their associated basins of attraction, because the objective is to match those features rather than just improve adherence to descent dynamics.
Lev: They validated this approach across three different datasets—crystal defects, pure crystals, and catalysts—showing that BPTT fine-tuning improves performance in all cases.
Kai: Qualitatively, they noted that the defect centers showed the most pronounced improvements in the resulting defect structures, which is a pretty telling detail.
Mira: So what this means for us is that this method lowers the data requirements for producing an effective domain specific machine learning interatomic potential by improving trajectory level behavior.
Lev: It's a powerful fine tuning tool, but they are clear about the limitations: BPTT fine-tuning isn't presented as a standalone learning paradigm; it needs to be part of a staged training pipeline.
Conclusion: Kai: So, looking at the full picture of "Worse Forces Can Lead to Better Structures: Trajectory-Level Fine-Tuning of Machine Learning Interatomic Potentials", it seems the authors are shifting our focus away from just chasing perfect force accuracy as a standalone goal.
Mira: They are arguing that for practical applications in materials design, what really matters is how well your model handles the entire process of structural relaxation, not just the initial force calculation accuracy.
Lev: The title itself points to this trade-off: you can accept forces that aren't perfectly accurate if you optimize the whole trajectory to get a better final structure.
Kai: It means we should start thinking about how we supervise our machine learning models based on the actual physical outcome they produce, not just by looking at local energy or force residuals.
Mira: This suggests a future where MLIPs are optimized not as perfect simulators of first principles forces, but as tools that learn to guide atomic configurations toward physically relevant final states.
Lev: It's a pragmatic move for high throughput studies because improved relaxation fidelity means you might not need the extremely expensive first principles DFT evaluations as often when studying specific structure classes.
Kai: So in simple terms, this paper suggests that trajectory level fine-tuning is a tool to bridge the gap between what our models learn and what we actually need for designing new materials.
Evan Dramko, Yizhi Zhu, Aleksander Krivokapic, Geoffroy Hautier, Thomas Reps
Department of Computer Science, Rice University · Department of Materials Science and NanoEngineering, Rice University · Rice Advanced Materials Institute, Rice University
cond-mat.mtrl-sci, cs.AI, cs.LG
Submitted: 2025-11-30
Updated: 2026-11-09
Comments: 10 main pages, total of 18 pages. 5 tables, 10 Figures
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 74/100
The gist: Accurate structural relaxation is critical for advanced materials design because traditional physics-derived methods are computationally expensive…
Key concepts
- Trajectory-Level Training Strategy
- This method involves unrolling a full atomic relaxation process into a sequence of steps (frames). Instead of training the model only on individual force predictions, the model is trained end-to-end by optimizing the final structure after many such steps. This treats the entire simulation trajectory as a single optimization problem.
- Coupled Optimization Levels
- The study explicitly links two different types of optimization: updating MLIP parameters (the force model) and updating atomic coordinates (the structural relaxation). By treating these updates as coupled processes, the framework learns how to adjust the force predictions specifically to guide the structure toward a desired outcome.
- Learning Structural Bias
- The fine-tuning process does not aim to perfectly reproduce first-principles forces. Instead, it learns a higher-level structural bias embedded in the training data. This results in an iterative map that preserves essential properties of the true potential energy surface, such as fixed points and their associated basins of attraction, which are crucial for accurate relaxation.
- Trade-off between Fidelity Aspects
- The paper demonstrates a trade-off: force fidelity (accuracy against first principles) decreases slightly while structural fidelity (accuracy of the final relaxed structure) increases substantially. This shows that conventional metrics might be incomplete, and optimizing for downstream performance can yield better scientific results than optimizing solely for local force accuracy.
Terminology
Summary
Accurate structural relaxation is critical for advanced materials design because traditional physics-derived methods are computationally expensive, motivating machine-learning interatomic potentials (MLIPs) which strive to faithfully reproduce first-principles computed forces The authors propose a fine-tuning method to be used on a pretrained MLIP in which they create a fully differentiable end-to-end simulation loop that optimizes the predicted final structures directly This method consistently improves performance across all evaluated pretrained models, resulting in an average of roughly 32% reduction in prediction error
The core methodology involves a trajectory-level training strategy
The authors adopt a trajectory-level training strategy where they unroll full relaxation trajectories and apply backpropagation through time (BPTT) to update model parameters based on the quality of the final relaxed structure rather than intermediate force errors This reframes MLIP training as a problem of optimizing the outcome of the relaxation process itself, allowing one data modality to supervise another Specifically, they define a frame as a single DFT/MLIP force-prediction step followed by a corresponding structural-update step The rollout is defined as Xb F = frame ◦ · · · ◦ frame [z] × k (X0) for k steps
The training signal is built through coupled optimization levels
The work centers on two distinct optimization processes: the MLIP training and the PES-level optimization of atomic coordinates during structural relaxation The key idea is to explicitly couple these two levels during training, treating parameter updates of the MLIP as MLIP training and coordinate updates as PES-level optimization They investigate various PES-level step-size controls, including a single learned scalar step size and a small neural network that predicts the step size dynamically during the relaxation trajectory
The gains come from learning structural bias rather than physical dynamics
The analysis suggests that BPTT fine-tuning is not improving the physical fidelity of force predictions nor optimizing the underlying relaxation dynamics in a conventional sense Instead, it learns a higher-level structural bias of the dataset by constructing a proxy iterative map that preserves key properties of the true PES, such as fixed points and their associated basins of attraction This is achieved because the objective is not to improve adherence to PES-level descent dynamics but to learn a map that matches fixed points and relevant basins-of-attraction to the PES
Validation across diverse structural domains confirms generalizability
The approach is validated across three datasets: crystal defects, pure crystals, and catalysts, showing that BPTT fine-tuning improves performance in all cases Qualitative analysis shows that the defect centers exhibit the most pronounced improvements in the resulting defect structures The results support a pragmatic interpretation that BPTT can reliably improve the trajectory-level behavior of MLIPs across many applied regimes, even if it does not solve the physics
The final recommendation is a staged training pipeline
The authors suggest treating BPTT as the final stage of a staged training pipeline: start from a pretrained, broadly applicable MLIP; refine it with supervised structure-to-force training on the target domain; then apply trajectory-level BPTT to shape the long-horizon relaxation This framing is supported by experimental results showing that BPTT fine-tuning is not an effective standalone learning paradigm but a powerful fine-tuning tool The benefit is not confined to extremely narrow tasks, as the pure crystal dataset exhibits substantial diversity spanning orders of magnitude in system size and covering a broad range of chemistries While BPTT does not solve the physics,
it can reliably improve trajectory-level behavior across many applied regimes The trade-off in high-throughput or computationally heavy studies of specific structure-classes is favorable because improved relaxation fidelity reduces the need for expensive DFT evaluations The authors conclude that BPTT fine-tuning lowers the data requirements for producing an effective domain-specific MLIP, addressing a common bottleneck in practical deployment
REFERENCE
arXiv:2512.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Fine-tuning MLIPs through an end-to-end simulation loop allows for optimization of atomic coordinates along the potential energy surface by tracking gradients through entire unrolled trajectories, which
reframes MLIP training as a problem of optimizing the outcome of the relaxation process itself.
This enables the model to learn ahigher-level structural bias
instead of just matching local force accuracy. -
The method can be used to produce improved relaxed structures for materials design tasks by leveraging trajectory-level signals, showing that
BPTT fine-tuning produces notably improved final configurations
even though it may degrade force prediction accuracy, as seen in Table 2. -
The system can be adapted to learn optimal PES-level step sizes by allowing BPTT to update the MLIP weights, which
learn a strong set of optimization controls
that are different from those typically used, potentially matching or outperforming domain-engineered optimizers.
Abstract
Machine learning interatomic potentials (MLIPs) are commonly trained and evaluated by their ability to reproduce first-principles energies and forces. Practical value, however, is often determined by downstream processes such as structural relaxation. Much of MLIP literature is motivated by the expectation that improving force accuracy will produce correspondingly better downstream behavior. Although this expectation is natural, it may not hold in practice due to model inaccuracies. Here, we demonstrate that force accuracy and relaxation accuracy can be driven in opposite directions. Starting from pretrained MLIPs, we differentiate through relaxation trajectories and optimize the the final structures. This intervention produces increases in force-prediction errors while substantially reducing structural errors. Importantly, this represents a trade-off between two aspects of physical fidelity: the learned forces become less faithful to first-principles forces, while the resulting relaxed structures become more faithful to first-principles predictions. Rather than observing a discrepancy between local quality and downstream performance, we demonstrate that this discrepancy can be induced through optimization. Our findings suggest that MLIPs need not be optimized as approximations of first-principles forces, but can instead learn task-specific vector fields that guide relaxation trajectories toward more accurate final structures. By propagating supervision through complete relaxation trajectories, our framework captures dependencies between successive predictions that are absent when structure--force pairs are treated independently. More broadly, these results demonstrate that conventional force metrics can provide an incomplete assessment of downstream performance and highlights the importance of evaluating scientific machine-learning models within computational workflows.
Sources
- Stabilizing Equilibrium Models by Jacobian Regularization
- Command A: An Enterprise-Ready Large Language Model
- ADAPT: Lightweight, Long-Range Machine Learning Force Fields Without Graphs
- Learning Inter-Atomic Potentials without Explicit Equivariance
- Positive concave deep equilibrium models
- Does RLHF Scale? Exploring the Impacts From Data, Model, and Method
- Transformers Discover Molecular Structure Without Graph Priors
- Explicit Inductive Bias for Transfer Learning with Convolutional Networks
- Proximal Policy Optimization Algorithms
- Interatomic Potential in a Simple Dense Neural Network Representation
- Secrets of RLHF in Large Language Models Part II: Reward Modeling
- MatterSim: A Deep Learning Atomistic Model Across Elements, Temperatures and Pressures
- Fine-Tuning Language Models from Human Preferences
Related papers
- Structural contribution to light-induced gap suppression in Ta 2 NiSe 5
- Anatomy of fast current-induced skyrmion motion in synthetic antiferromagnets
- Magnetic control of an exciton-polariton condensate in a van der Waals magnet
- High-pressure elastic properties of GeO2 polymorphs up to 120 GPa
- Decoupling Strain-Rate Sensitivity and Deformation Length Scale Effects in Neutron-Irradiated Tungsten: A Coupled Nano-Indentation, HR-EBSD and Crystal Plasticity Study
- Magnetism modulation in Co3Sn2S2 by current-assisted domain wall motion