Learning Thermal Response Forces: A Method for Extending the Thermodynamic Transferability of Coarse-Grained Models via Machine-Learning

Patrick G. Sahrmann, Benjamin Nebgen, Kipton Barros, Brenden W. HamiltonarXiv 2026 · DOI · arXiv

TL;DR

Researchers have developed a novel way to make "coarse-grained" molecular models work across a wide range of temperatures using data from just a single temperature point. By training machine learning models on "thermal response forces"—the derivatives of the potential with respect to temperature—they can capture how a system's underlying physics shifts as it heats up or cools down, significantly improving the accuracy and predictive power of simplified simulations.

What Problem Does This Paper Address?

In the world of molecular simulation, "coarse-graining" (CG) is a common trick to save time. Instead of simulating every single atom, you group clusters of atoms into single "beads." This makes simulations orders of magnitude faster, but it comes with a catch: transferability.

Unlike the fundamental forces between atoms, the effective forces in a CG model (derived from the "Potential of Mean Force" or PMF) are actually a form of free energy. Because free energy depends on temperature, a CG model trained to simulate water at room temperature usually fails miserably if you try to use it to simulate boiling water or ice. Traditionally, fixing this required running expensive "fine-grained" simulations at many different temperatures to "teach" the model how to adapt. This paper seeks a more data-efficient way to bake temperature dependence directly into the model's DNA.

Methods

The authors propose a theoretical breakthrough: instead of just learning the forces at one temperature, the model should learn the thermal response forces.

  1. Taylor Expansion: They treat the CG free energy as a Taylor series around a reference temperature. The first-order correction relates to entropy, and the second-order correction relates to heat capacity.
  2. Force-Matching Derivatives: They derive exact statistical mechanical identities that allow them to calculate "entropic forces" and "heat capacity forces" from a single fine-grained simulation.
  3. Machine Learning (HIP-NN-TS): They use a specialized neural network architecture (Hierarchically Interacting Particle Neural Network with Tensor Sensitivity) to learn these different force components simultaneously.
  4. Model Variants: They compared a standard CG model (0th order) against TCG1 (includes entropy) and TCG2 (includes entropy and heat capacity).

Datasets

The study focused on a classic benchmark: Liquid Water.

  • Fine-grained model: SPC/Fw water (a flexible, all-atom model).
  • Mapping: Each water molecule was mapped to a single CG site at its center of mass.
  • Training Points: Models were trained using data from simulations at 250 K, 300 K, and 350 K.
  • Evaluation: The models were tested across a massive temperature range, from 70 K to 1000 K, to see how well they could extrapolate.

Key Insights and Ideas

  • Entropy has "Forces": We usually think of entropy as a global property of a system. This paper highlights that there are local "entropic forces" acting on CG beads that can be calculated and used as training targets for machine learning.
  • Local vs. Global Learning: The authors argue that training on these local response forces is much more efficient than trying to match global properties like internal energy. It provides $3N$ (where $N$ is the number of atoms) training signals per configuration rather than just one.
  • Predicting the "Speed-Up": CG simulations naturally run "faster" than real-time because they lack the friction of the missing atoms. The authors found a linear relationship between the temperature and the log-ratio of diffusion constants, allowing them to mathematically rescale CG time to match real-world time accurately.
  • Asymmetric Transferability: The Taylor series approach works better when moving to higher temperatures ($\Delta T > 0$) than lower ones. This is because the PMF varies more smoothly as things get hotter and more disordered, whereas cooling down often leads to complex structural changes (like freezing).

Key Results

  • Structural Accuracy: The TCG2 model (second-order) maintained highly accurate Radial Distribution Functions (RDFs)—a measure of how molecules pack together—across hundreds of degrees of Kelvin, whereas the standard CG model failed quickly outside its training point.
  • Dynamic Fidelity: The TCG models successfully captured non-Arrhenius behavior in diffusion. In simpler terms, they correctly predicted that water's "runniness" doesn't change linearly with temperature, a feature usually lost in basic CG models.
  • Extrapolation: The models trained at 300 K remained physically interpretable and reasonably accurate even when extrapolated to 700 K, a testament to the robustness of the thermal response framework.

Limitations and Future Work

  • The Taylor Series Limit: The authors noted that the Taylor expansion eventually breaks down, especially at very low temperatures where the assumption that the PMF "varies slowly" no longer holds.
  • Missing Friction: While the model captures the equilibrium forces perfectly, it still lacks the explicit "noise and memory" terms (from the Mori-Zwanzig formalism) that represent the missing atoms' friction. The authors used a time-rescaling post-processing step to fix this, but building these terms directly into the ML model is a logical next step.
  • Phase Transitions: The current method was tested on liquid water. It remains to be seen how well these thermal response forces can handle "Sobolev training" across phase boundaries (e.g., from liquid to gas).

Summary updated

Abstract

Machine-learned (ML) coarse-grained (CG) models are a promising tool for significantly enhancing the efficiency of molecular simulations by systematically removing degrees of freedom while retaining fidelity to the underlying fine-grained model. The CG potential of mean force (PMF) is inherently dependent on thermodynamic conditions and, hence, a CG force-field (FF) which is trained at one thermodynamic state point is not necessarily accurate at another. We propose, in this work, a novel and data-efficient means of learning temperature dependence into ML CG force-fields via training on the thermal response forces of the PMF. We demonstrate how incorporating these terms into ML CG FFs confers significantly improved transferability for CG water models and demonstrate how this transferability enables accurate and predictive CG dynamics.
BibTeX
@misc{sahrmann2026learning,
  author = {Patrick G. Sahrmann and Benjamin Nebgen and Kipton Barros and Brenden W. Hamilton},
  title = {{Learning Thermal Response Forces: A Method for Extending the Thermodynamic Transferability of Coarse-Grained Models via Machine-Learning}},
  year = {2026},
  eprint = {2602.23198},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2602.23198},
}

Loading related content...

Figure