Improving SLICES crystal representation through CHGNet integration and parameter tuning

In the rapidly evolving field of materials science, the inverse design of crystals has become increasingly significant with the widespread adoption of generative models. However, current crystal representation methods face persistent challenges in maintaining both invertibility and invariance. To bridge this gap, we present an enhanced approach to the simplified line-input crystal-encoding system (SLICES) representation by incorporating the Crystal Hamiltonian Graph Neural Network (CHGNet) machine learning force field model. Comprehensive evaluations across multiple datasets (MP-20, MP21-40, and MOF) using different neighbor recognition algorithms (EconNN and CrystalNN) show that our approach outperforms the original in reconstructing structures with fewer than 20 atoms per unit cell, achieving up to a 1.34% improvement in reconstruction rate (from 92.55% to 93.89%). Furthermore, through systematic optimization of key parameters, including bond scaling, Δx, and lower bounds of lattice scaling, we achieved enhanced reconstruction performance. Our research represents a step forward in developing more efficient solutions for crystal inverse design utilizing the SLICES representation framework.