Cost-Efficient Fine-Tuning of Neural Network Potentials for Modeling Barocaloric Phase Transitions ✛ ✦
Context
To achieve cooling without harmful greenhouse gases, solid-state caloric materials are being actively studied, where a pressure-induced phase transition causes a significant change in entropy—a concept introduced by Ludwig Boltzmann. Understanding and predicting such phase transformations is critical, and here machine-learned interatomic potentials (MLIPs) come to the rescue, capable of modeling the dynamics of hundreds or thousands of atoms with an accuracy comparable to quantum theory methods. However, generating training data via density functional theory (DFT) remains computationally expensive, motivating the search for strategies to reduce the required data volume without sacrificing accuracy.
Methods
The scientists used a compound containing hydrogen, oxygen, and ammonium—ammonium sulfate. A set of configurations was collected from DFT calculations with the r2SCAN functional, a method rooted in the quantum mechanics of Max Planck. Models based on the MACE architecture were trained in three ways: from scratch, and via naive (NFT) and multi-head with repetition (MHFT) fine-tuning of the pre-trained base model MACE-MPA-0. Effectiveness was evaluated by the ability to reproduce the phase transition in molecular dynamics simulations at temperatures from 100 to 400 K.
Results
Models trained from scratch required at least 50 DFT configurations to accurately describe the transition, while the fine-tuned versions successfully reproduced lattice parameter jumps even with 5–10 training examples, despite the original base model's complete inability. Quantitatively, the force error for the 10-sample model was about 37 meV/Å. Moreover, it turned out that without dispersion corrections in DFT, describing the transition is impossible—consistent with the key role of the hydrogen bond network, whose rearrangement underlies the second law of thermodynamics—the barocaloric effect itself is tied to changes in entropy. Both fine-tuning strategies yielded comparable results, although MHFT showed greater robustness on out-of-training configurations.
Implications
The demonstrated extreme data efficiency means that for new materials, just a handful of DFT calculations suffice, freeing up computational resources for more accurate methods like hybrid functionals. This brings machine learning and quantum thermodynamics of phase transitions closer together, enabling faster discovery of efficient refrigerants.
Future development
In the future, this approach could be extended to other classes of caloric materials—magnetocaloric, electrocaloric—which will require incorporating additional degrees of freedom. Moreover, combining with accelerated sampling methods will allow the study of phase transitions that demand orders of magnitude more simulation time.
Impact
The results directly impact the accelerated search for new solid-state refrigerants and the development of machine learning methods in materials science, shortening the path from quantum calculations to macroscopic properties.
Next steps
Next steps include testing the fine-tuning protocols on a wider range of substances and integrating experimental data to validate predictions.
Key open problems
The work touches on the challenge of accurately describing van der Waals interactions in DFT, which is critical for molecular crystals and hydrogen-bond-driven phase transitions, and highlights the need for further development of dispersion corrections.
The knowledge machine's view
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
This study explores minimizing the training set for machine-learned interatomic potentials (MLIPs) when modeling the barocaloric effect in ammonium sulfate. Three training strategies for the MACE architecture are compared: training from scratch, naive fine-tuning (NFT), and multi-head fine-tuning with repetition (MHFT) of the pretrained MACE-MPA-0 model. It is shown that fine-tuning can reproduce the phase transition with just 5–10 DFT configurations, instead of 50+ required for training from scratch, with dispersion corrections playing a key role by capturing the rearrangement of the hydrogen-bond network.
Where its strength is
The strength of this work lies in demonstrating extreme data efficiency: fine-tuning a universal model on a tiny set of specific configurations captures a phase transition that the base model completely misses. This paves the way for rapid screening of caloric materials with minimal computational cost.
Where it could go next
- It would be interesting to see how a similar fine-tuning approach for universal MLIPs works for other material classes with strong anharmonicity and hydrogen-bond rearrangements, such as hybrid perovskites or metal-organic frameworks, where dispersion interactions are also important. In our archive, there is a related work where a neural network learns from the errors of quantum chemistry, which might offer insights into improving transfer of pretrained models to new chemical spaces.based on: How a Neural Network Learns from Quantum Chemistry's Mistakes arXiv:2604.24607
- Another avenue is to extend the multi-head fine-tuning strategy to describe not only structural but also electronic properties, such as predicting polarization or dielectric response, which is crucial for caloric materials with ferroelectric transitions. The archive contains an analysis where machine learning accelerates molecular calculations using warm starts, which may provide useful ideas for adapting models to new properties.based on: Warm Start for a Quantum Engine: How Machine Learning Accelerates Molecular Calculations arXiv:2607.09304
- A possible direction is to test the robustness of fine-tuned potentials when simulating far from the training set, for example at high pressures or temperatures where new phases may emerge. There is a related work on AI-driven discovery of chameleon materials that discusses approaches to exploring phase spaces, which could be useful for planning active learning.based on: How AI hunts for chameleon materials arXiv:2607.03578
Worth trying
- Test active learning: starting with 5 configurations, iteratively add points with the highest prediction uncertainty to further reduce the dataset.
- Investigate transferability of fine-tuning to other caloric materials, such as halide perovskites, using the same DFT protocol with dispersion corrections.
- Compare MHFT with alternative adaptation methods like LoRA for the same task to evaluate computational efficiency and accuracy.
Why this may matter
This work shows that fine-tuning universal MLIPs can drastically lower the barrier to modeling complex phase transitions, enabling high-throughput screening of solid-state refrigerants. It brings practical application of barocaloric materials in eco-friendly cooling closer to reality.
Nearby works in our archive
- Neural Network Learns to Read Molecules by Their Carbon Signaturesimilarity 0.61
- Proteins on a Quantum Ropesimilarity 0.61
- Neutrino Rebellion Decoder: How AI Reads Chaos in Stellar Coressimilarity 0.61
- Memory for a quantum computer at minus 271 degreessimilarity 0.60
- How a Neural Network Learns from Quantum Chemistry's Mistakessimilarity 0.60
- Quantum Dance in a Protective Cocoonsimilarity 0.59
- How AI hunts for chameleon materialssimilarity 0.59
- Warm Start for a Quantum Engine: How Machine Learning Accelerates Molecular Calculationssimilarity 0.59
The closest work in the archive sits at 0.61 by meaning — the usual level for our corpus.
Original abstract
as written by the authors arXiv:2606.25742
Solid-state cooling based on the barocaloric (BC) effect has emerged as promising environmentally friendly prospective alternative to conventional vapor-compression refrigeration. The search for suitable BC materials relies on efficient atomistic simulations of their phase behavior. Machine-learned interatomic potentials (MLIPs) enable such simulations at near density functional theory (DFT) accuracy, but generating the required DFT training data remains computationally demanding, which motivates development of strategies that reduce the amount of data needed to train accurate models. In this work, we use a prototypical BC material, ammonium sulfate, as a model system and investigate how small a training set can be while still reproducing the temperature-driven structural phase transformation that underlies its BC response. We train a series of MLIPs based on the MACE architecture using three strategies: training from scratch, and naive- and multihead replay fine-tuning of the MACE-MPA-0 foundation model. These strategies are evaluated on their ability to reproduce the phase transformation of ammonium sulfate in molecular dynamics simulations across a range of training-set sizes. We find that, while the MACE-MPA-0 foundation model itself fails to reproduce the transformation, and models trained from scratch break down for small datasets, fine-tuned models reproduce the transformation using as few as 5 to 10 60-atom DFT configurations. Both fine-tuning protocols yield similarly accurate results for ammonium sulfate, but we also find some indications that multihead replay fine-tuning is more robust on configurations outside the fine-tuning domain. Exploiting this data efficiency, we further show that models can be trained on small datasets and the hybrid-DFT level, and that some form of inclusion of dispersion correction is necessary to describe the phase behavior correctly.
Just five quantum-mechanical calculations — the price the authors 'paid' for the neural network to 'understand' the complex structural transition, even though without extra training it was off by a mile.
Key numbers
- minimum number of training configurations: 5
- shift of transition temperature in the model: ~60 K
- root-mean-square force error (NFT, 10 samples): 37 meV/Å
- supercell size in MD: 1080 atoms
- experimental transition temperature: 219 K