🔬 materials science

Barocaloric phase transformation from data efficient fine-tuning of machine learned interatomic potentials

This study demonstrates that fine-tuning the MACE-MPA-0 foundation model on as few as 5 to 10 DFT configurations enables accurate machine-learned interatomic potentials to reproduce the barocaloric phase transformation of ammonium sulfate, overcoming the data inefficiency of training from scratch while highlighting the necessity of dispersion corrections and hybrid-DFT levels for correct phase behavior description.

Original authors: Ludwig Hedin, Johan Klarbring

Published 2026-06-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ludwig Hedin, Johan Klarbring

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Finding Better Coolants

Imagine we are trying to invent a new kind of air conditioner. Instead of using harmful gases (like the ones in your fridge that hurt the planet), scientists want to use solid materials that get hot or cold just by being squeezed or released. This is called the "barocaloric" effect.

To find the perfect solid material, scientists need to run computer simulations to see how the atoms inside the material move and change shape when squeezed. However, these simulations are like trying to predict the weather: they are incredibly difficult and require massive computing power.

The Problem: The "Training" Bottleneck

To make these simulations accurate, scientists use "Machine Learned Interatomic Potentials" (MLIPs). Think of an MLIP as a super-smart apprentice who learns the rules of how atoms behave.

The Old Way (Training from Scratch): To teach this apprentice, you usually have to show them thousands of examples of atoms interacting, calculated by a very slow, super-accurate computer program (called DFT). It's like trying to teach a child to drive by making them sit through 10,000 hours of driving school before they are allowed to touch the wheel. It takes too long and costs too much money.
The New Idea (Foundation Models): Recently, scientists created a "Foundation Model." This is like a world-class driving instructor who has already learned how to drive in every possible condition (rain, snow, city, highway) by studying a massive library of data.

The question this paper asks is: If we have this world-class instructor, how many extra lessons do we need to teach them to drive a specific, weird car (Ammonium Sulfate) perfectly?

The Experiment: Teaching the Apprentice

The researchers picked a specific material called Ammonium Sulfate (a type of salt) because it is a known "cooling champion" that changes its structure when squeezed. They wanted to see how little data they could use to train their AI models to predict this change correctly.

They tried three different teaching strategies:

Start from Scratch: Teach the apprentice from zero, showing them only a few examples.
Naive Fine-Tuning (NFT): Take the world-class instructor and just keep teaching them on the new car, ignoring their old training.
Multihead Replay Fine-Tuning (MHFT): Take the world-class instructor, teach them the new car, but also occasionally remind them of their old training (the "replay") so they don't forget how to drive in general.

The Results: Less Data, Better Results

The findings were surprising and very efficient:

The "From Scratch" Failure: When they tried to teach the apprentice from scratch with very few examples (like 10 or 25), the apprentice got confused and failed to predict the material's behavior. They needed about 50 examples just to get it "okay."
The "Fine-Tuning" Success: When they used the pre-trained world-class instructor, the results were amazing.
- They only needed 5 to 10 examples (out of thousands) to get the model to work perfectly.
- It's as if the instructor already knew 99% of the rules, and they only needed to learn 1 or 2 specific quirks of this new car to drive it flawlessly.

Both "Fine-Tuning" methods (NFT and MHFT) worked well for this specific material. However, the Multihead Replay (MHFT) method was slightly more robust. It was better at remembering the general rules of physics when asked about things it hadn't seen before, whereas the "Naive" method started to "forget" its general knowledge as it focused too hard on the specific material.

A Crucial Detail: The "Glue" (Dispersion)

The researchers also discovered something important about the "ingredients" used to train the model.

They found that to accurately describe how the atoms in Ammonium Sulfate stick together, the computer model needed to account for a specific type of weak attraction called dispersion (think of it as the "molecular glue" or Velcro that holds things together).
If they didn't include this "glue" in their training data, the model failed, even if it was fine-tuned.
Interestingly, the standard "world-class instructor" (MACE-MPA-0) failed on its own because it was trained on data that didn't include this glue. But once they fine-tuned it with data that did include the glue, it worked perfectly.

The Bottom Line

This paper proves that you don't need a massive library of data to train a super-accurate AI model for specific materials. By using a "world-class instructor" (a foundation model) and giving it just a handful of specific examples (5 to 10), you can get results that are just as good as training from scratch with hundreds of examples.

This makes the search for new, eco-friendly cooling materials much faster and cheaper, because scientists can now use more advanced, accurate computer methods without waiting weeks for the data to be generated.

Technical Summary: Barocaloric Phase Transformation from Data Efficient Fine-Tuning of Machine Learned Interatomic Potentials

Problem Statement
Solid-state cooling based on the barocaloric (BC) effect is a promising alternative to conventional vapor-compression refrigeration, which relies on potent greenhouse gases. Identifying suitable BC materials requires characterizing pressure-driven structural phase transformations. While atomistic simulations are essential for this, they face a trade-off: classical molecular dynamics (MD) is efficient but lacks accuracy, whereas Density Functional Theory (DFT) is accurate but computationally prohibitive for the system sizes and timescales required to observe phase transitions. Machine-learned interatomic potentials (MLIPs) offer a middle ground, providing near-DFT accuracy at a fraction of the cost. However, training accurate MLIPs typically requires large, computationally expensive DFT datasets. This paper addresses the challenge of reducing the data requirements for training MLIPs to accurately reproduce the temperature-driven phase transformation in barocaloric materials, specifically ammonium sulfate.

Methodology
The authors use ammonium sulfate as a prototypical model system, which undergoes a first-order phase transformation from a high-temperature centrosymmetric orthorhombic phase ($Pnam$) to a low-temperature polar orthorhombic phase ( $Pna2_1$ ) at approximately 219 K.

The study employs the MACE (Machine Learning Atomic Cluster Expansion) architecture and compares three training strategies:

Training from scratch: Initializing MACE models with random weights and training on datasets of varying sizes (5 to 500 configurations).
Naive Fine-Tuning (NFT): Starting from the pre-trained MACE-MPA-0 foundation model weights and continuing training on a custom, smaller dataset.
Multihead Replay Fine-Tuning (MHFT): Fine-tuning the foundation model on the custom dataset while simultaneously training a "replay head" on a subset of the original pre-training data (containing H, S, N, O elements) to prevent catastrophic forgetting of the general chemical knowledge.

The training data consisted of DFT configurations generated using the r2SCAN meta-GGA functional. To test data efficiency, the authors constructed subsets of 60-atom structures ranging from 5 to 500 samples using the DIRECT (Dimensionality-Reduced Encoded Clusters with sTratified) sampling workflow. These models were evaluated on a hold-out test set of 360-atom structures. Additionally, the authors investigated the use of hybrid DFT (PBE0) with and without D3 dispersion corrections to assess the impact of higher-level theory and dispersion interactions on the phase behavior.

Key Results

Foundation Model Limitations: The pre-trained MACE-MPA-0 foundation model failed to reproduce the correct qualitative behavior of the phase transformation, predicting a transition to the wrong state.
Data Efficiency of Fine-Tuning: Fine-tuned models (both NFT and MHFT) successfully reproduced the phase transformation using as few as 5 to 10 60-atom DFT configurations. In contrast, models trained from scratch required approximately 50 configurations to achieve a reasonable description and failed completely with 10 or fewer configurations.
Comparison of Fine-Tuning Strategies:
- For the specific ammonium sulfate task, both NFT and MHFT yielded similarly accurate results for small datasets (5–100 samples).
- As dataset size increased, NFT models showed lower errors on the target task, approaching the performance of scratch-trained models. However, MHFT models reached an accuracy floor around 200 samples, likely due to the constraint of the replay head preventing full specialization to the target data.
- Out-of-Domain Generalization: MHFT demonstrated superior robustness on out-of-domain configurations (chemically similar but not ammonium sulfate). NFT models exhibited "catastrophic forgetting," where increasing the training set size degraded their ability to generalize to unseen materials, whereas MHFT maintained stability.
Role of Dispersion Corrections: Models trained on hybrid PBE0 data without dispersion corrections failed to reproduce the phase transformation, similar to the foundation model. However, adding D3 dispersion corrections (PBE0+D3) allowed the models to correctly describe the phase behavior. This suggests that dispersion interactions, driven by hydrogen-bond network reorganization, are critical for this system. The r2SCAN functional used for the primary dataset likely captured these interactions through its exchange description, obviating the need for explicit dispersion corrections in that specific case.

Significance and Claims
The paper claims that fine-tuning foundation models is a highly data-efficient strategy for constructing MLIPs capable of capturing complex phase transformations. The primary contribution is demonstrating that accurate models for specific materials can be generated from extremely small datasets (5–10 configurations) by leveraging pre-trained foundation models, a feat not achievable by training from scratch.

The authors emphasize that this data efficiency enables the use of more computationally expensive, higher-accuracy functionals (such as hybrid DFT) for training data generation, which would otherwise be infeasible for large datasets. While the study focuses on ammonium sulfate, the results suggest a practical route for the computational screening of a larger number of potential barocaloric materials. The authors modestly note that while the phase transformation in ammonium sulfate is resolvable within ~100 ps MD trajectories, other materials may require longer timescales or enhanced sampling techniques. The study does not claim universal applicability across all materials without further validation but highlights the potential of this workflow for targeted material discovery.