Refining catalyst–adsorbate interatomic potentials with transfer learning in ænet-PyTorch From optimizing catalyst interfaces to extending molecular dynamics (MD) simulations, linking broad chemical knowledge to specific adsorbate systems often poses challenges in materials research. While large-scale data repositories can help, constructing accurate machine learning potentials (MLPs) for adsorbate-catalyst complexes still requires significant computational resources, especially if only a small custom data set is available. A recent paper by An Niza El Aisnada and coauthors proposes a transfer learning strategy to build stable MLPs under tight data constraints, particularly for catalyst–adsorbate systems. Leveraging the Open Catalyst 2020 (OC20) database—a substantial collection of diverse catalyst configurations—they pretrain MLPs on carefully selected OC20 subsets. By transferring the pretrained models to a smaller target data set (only a few hundred ab initio references), they achieve robust energy and force predictions. Notably, these transfer-learned MLPs remain stable for hundreds of picoseconds of MD simulation on Cu–Au/water cluster systems, whereas models trained only on limited local data fail much sooner. They explore two main approaches for selecting relevant subsets from OC20: (1) random sampling to mirror the original database broadly, and (2) filtering by chemical environment (for example, focusing on Cu–Au). The pretrained MLPs, once transferred, exhibit significant improvements in force prediction and MD stability—even though raw RMSE metrics in smaller data sets do not always reflect such gains. A key component of their workflow is the “ænet-PyTorch” framework. Originally, the Atomic Energy Network (ænet) was a C/Fortran toolkit for ANN-based MLP construction. In this updated PyTorch extension, parallelization and GPU acceleration are harnessed for efficient training, allowing the incorporation of both reference energies and forces. Through transfer learning, a user can import a pretrained model (from large data sets), then fine-tune it on domain-specific references to achieve both accuracy and scalability. Beyond a simple methods comparison, the authors emphasize pragmatic insights—such as the importance of CV-limited data curation, the synergy of domain-focused subset selection (e.g., focusing on Cu–Au to boost transfer success), and the pitfalls of relying on single scalar metrics like RMSE. They illustrate how data set sizes and neural network hyperparameters (for balancing energy vs. forces) drive generalizability in practice. Paper: pubs.acs.org/doi/full/10.102…