Barocaloric phase transformation from data efficient fine-tuning of machine learned interatomic potentials
Ludwig Hedin, Johan Klarbring · 2026
Read Barocaloric phase transformation from data efficient fine-tuning of machine learned interatomic potentials by Ludwig Hedin, Johan Klarbring (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-25742.txt.
Introduction
Squeeze a barocaloric material and it warms; release the pressure and it cools, a solid state route to refrigeration without greenhouse gases. Finding good candidates needs atomistic simulation, and machine learned potentials can mimic quantum accuracy if fed enough quantum calculations, which are expensive. Ludwig Hedin and Johan Klarbring ask in 2026 how little training data suffices, using ammonium sulfate, whose temperature driven structural change powers its cooling effect. Three training strategies compete on the MACE architecture across shrinking datasets.
The untuned foundation model failed to reproduce the phase change at all, and models trained from scratch broke down on small datasets. Fine tuned models, by either of two protocols, captured the transformation from as few as five to ten small quantum configurations. A dispersion correction proved necessary for correct phase behaviour, and the data frugality even allows training at a costlier hybrid quantum level.
Catalog record
| CREATOR | Ludwig Hedin, Johan Klarbring |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 36 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2606.25742
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Barocaloric phase transformation from data efficient fine-tuning of machine learned interatomic potentials Ludwig Hedin, Johan Klarbring Abstract 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 barocaloric 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 barocaloric 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. I Introduction The hydrofluorocarbon (HFC) refrigerants used in conventional vapor-compression cooling are potent greenhouse gases, with global warming potentials thousands of times greater than that of \mathrm{CO_{2}} 2 , and with refrigeration consuming a large and growing share of global energy 3 , more sustainable alternatives are needed. Solid-state cooling based on caloric materials is one such alternative, exploiting external field driven phase transformations that produce substantial temperature changes under adiabatic conditions. While magnetocaloric compounds have historically been the most extensively studied 9 ; 11 ; 42 , barocaloric (BC) materials, where the transformation is driven by hydrostatic pressure, have recently attracted strong interest. Indeed, large barocaloric effects have recently been demonstrated in a range of material classes, including, for example, colossal effects in plastic crystals of neopentyl glycol (NPG) 34 ; 33 . The barocaloric response of a material is governed by a phase transformation, and identifying promising candidates therefore requires characterizing this transformation. Atomistic simulations have become an indispensable tool for this task, allowing the relevant phase behavior to be characterized without recourse to experiment. Classical molecular dynamics (MD) enables efficient large-scale simulations but is often hindered by the limited accuracy of the underlying interatomic potential. In contrast, density functional theory (DFT) provides accurate calculations, yet remains limited to relatively small systems of \sim 10^{3} atoms and short time scales, due to its inherent computational complexity 49 ; 41 . Machine learned interatomic potentials (MLIPs) promise to circumvent these issues 27 ; 13 ; 22 ; 48 . Indeed, modern MLIP architectures, when trained on well-constructed DFT-datasets of energies, forces and stress, retain near DFT-accuracy, while reducing computational cost by orders-of-magnitude 49 . However, despite the success of MLIPs, there still remains significant ambiguity regarding how to optimize training strategies. Moreover, obtaining high-quality training data from DFT computations demands substantial computational resources. Consequently, it becomes of central importance to employ training strategies aimed at reducing the amount of required training data without sacrificing accuracy. So called foundation, or universal, MLIPs, trained on large chemically and structurally diverse datasets have gained prominence over the last few years 46 . These models offer qualitative accuracy and stability in MD simulations over a wide range of chemistries and structures, but do not offer the same quantitative accuracy for specific material systems as custom-trained MLIPs on targeted datasets 5 . Fine-tuning these pre-trained foundation models, rather than training an MLIP from scratch, has emerged as a promising route to data efficient training, and recent work has established that it improves accuracy across a range of MLIP architectures and target properties 20 ; 26 ; 18 ; 44 ; 36 . In this work, we build on these insights and use the prototypical BC material ammonium sulfate 35 as a model system. We ask specifically, how small can we make the training set, while still faithfully reproducing the structural phase transformation which underlies the functionality of barocaloric materials? We train a series of MLIPs based on the MACE architecture 6 using different strategies and dataset sizes. Specifically, we investigate three training strategies: training MACE models from scratch, as well as fine-tuning a MACE foundation model using two distinct approaches: multihead replay fine-tuning and naive fine-tuning. These models are then evaluated on their ability to reproduce the temperature-driven first-order phase transformation in ammonium sulfate from a series of MD simulations. We find that fine-tuned models reproduce the qualitative behavior of the transformation down to training sets of just five 60-atom DFT configurations. We further exploit this data efficiency and showcase that models can be efficiently constructed at beyond semi-local DFT accuracy by training models on small datasets of hybrid PBE0+D3 calculations. We find that some level of dispersion correction is essential in describing the phase behavior correctly, consistent with the transformation being driven by a reorganization of the hydrogen-bond network. II Methods II.1 Ammonium sulfate phases The barocaloric effects in ammonium sulfate, discovered by Lloveras et al. 35 in 2015, arise from a pressure-induced structural phase transformation at \sim 219\text{\,}\mathrm{K} . In the high-temperature regime, ammonium sulfate stabilizes in a centrosymmetric orthorhombic structure ( Pnam ), while it adopts an orthorhombic polar structure ( Pna2_{1} ) when cooled below the phase transformation temperature, see Fig. 1 . As the barocaloric effects in ammonium sulfate have been extensively studied 35 ; 39 ; 47 , it serves as an ideal model system for this study. \begin{overpic}[figures/LT.png] \put(3.0,40.0){(a)} \end{overpic} \begin{overpic}[figures/HT.png] \put(3.0,40.0){(b)} \end{overpic} Figure 1 : Unit cell of ammonium sulfate in (a) the low-temperature Pna2_{1} phase, and (b) the high-temperature Pnam phase. H, N, O and S atoms are shown in white, blue, red, and yellow, respectively. II.2 Dataset generation We initially generated two large and diverse datasets of relevant configurations. The first set contains 360-atom structures generated by running MD using the VASP machine learning force field (MLFF) implementation 25 ; 23 ; 24 , where relevant snapshots are picked out and added to the dataset based on an on-the-fly evaluation of the model error. Separate runs were performed for a range of temperatures and pressures in both the low-temperature (LT) and high-temperature (HT) phases. This dataset contains 1385 structures, out of which 139 structures were set aside to construct a hold-out test set. A second set with 60-atom structures were obtained by running MD with the MACE-MPA-0 38 foundation model at a set of temperatures and pressures, initialized in both the LT and HT phases, and randomly picking out a set of temporally well-separated structures. This dataset contains