Developing transferable coarse-grained (CG) models is a major challenge in molecular simulations, as conventional potentials are often state-point and system composition dependent. While neural network (NN) potentials offer a promising path to transferability, their development is hampered by training instabilities and the complex, often heuristic, selection of model parameters. This work presents a systematic methodology to address these issues, focusing on the rational selection of descriptors for NN-based CG potentials. We propose a systematic workflow for parameterizing Behler–Parrinello symmetry functions (G₂) by analyzing their resolving power and response to changes in system conditions, such as concentration. This allows for the selection of an optimized set of descriptors that captures essential structural features, handles short-range repulsions efficiently, and includes descriptors sensitive to the thermodynamic state to ensure transferability. Furthermore, we adopt a network extension technique that enables iterative improvement of model accuracy by expanding the network architecture and descriptor set without discarding prior training. The methodology is demonstrated on a CG model of methanol–water mixtures, with the NN trained to reproduce radial distribution functions derived from atomistic simulations across a wide range of methanol concentrations (10%–100%). The resulting NN potential, built using the systematic approach, demonstrates significantly improved accuracy and transferability, particularly at high concentrations, outperforming models developed previously with empirically chosen parameters. Our findings provide practical guidelines and a robust workflow for developing accurate and transferable NN potentials for CG simulations, paving the way for more reliable mesoscale modeling of complex systems.

We propose a method of bottom-up coarse-graining, in which interactions within a coarse-grained model are determined by an artificial neural network trained on structural data obtained from multiple atomistic simulations. The method uses ideas of the inverse Monte Carlo approach, relating changes in the neural network weights with changes in average structural properties, such as radial distribution functions. As a proof of concept, we demonstrate the method on a system interacting by a Lennard-Jones potential modeled by a simple linear network and a single-site coarse-grained model of methanol-water solutions. In the latter case, we implement a nonlinear neural network with intermediate layers trained by atomistic simulations carried out at different methanol concentrations. We show that such a network acts as a transferable potential at the coarse-grained resolution for a wide range of methanol concentrations, including those not included in the training set.