From Algebraic Completeness to Orbit Separation: Recursive Separating Invariants for Global Neural Network Potential Energy Surfaces

Abstract

Permutationally invariant polynomials (PIPs) provide a systematically improvable, symmetry-preserving representation for global potential energy surfaces (PESs), but their high computational cost as neural network descriptors limits applications to large molecular systems. Fundamental invariants (FIs), which form an algebraically complete generating set, substantially reduce the number of descriptors and enable neural network PESs for larger systems. However, PES evaluation remains computationally expensive as molecular size and complexity increase. Here, we introduce recursive separating invariants (RSIs), shifting the descriptor design principle from algebraic completeness to orbit separation. RSIs are constructed systematically through correlation order and recursion depth, providing a more computationally efficient representation. Benchmarks on reactive and nonreactive systems demonstrate that RSI-based neural network (RSI-NN) PESs achieve approximately an order-of-magnitude faster evaluation than their FI-NN counterparts, with system-dependent prediction accuracy that is comparable in several cases. These results establish orbit separation as a promising principle for developing efficient symmetry-preserving descriptors for global neural network PESs.