Exact CHY Integrand Construction Using Combinatorial Neural Networks and Discrete Optimization

Abstract: We present a novel approach to the inverse CHY problem-constructing integrands from desired pole structures-by establishing a correspondence between CHY combinatorics and Combinatorial Neural Networks (CNNs). Our key innovation is the generalized pole degree $K(s_A)$, which is additive under integrand multiplication and satisfies hierarchical recursion relations. These recursions naturally map to CNN message-passing architectures, enabling discrete automatic differentiation that preserves exact integer constraints throughout the computation. Unlike conventional machine learning approaches that yield numerical approximations, our method leverages the intrinsic combinatorial structure of the CHY formalism to guarantee exact algebraic solutions. This work demonstrates how specialized neural architectures can solve fundamental physics problems requiring strict mathematical constraints, providing a modest step toward bridging combinatorics, machine learning, and scattering amplitude theory.