CBANN: Causal neural surrogates with built-in constraints for predictive modeling and immeasurable latent recovery
Year: 2026
Type: article
Abstract: Industrial plants often operate with sparse, delayed, or end-of-process measurements, which makes accurate data-driven modeling and control challenging. Purely data-driven predictors such as multi-layer perceptrons (MLPs) rely on statistical correlations and may extrapolate poorly when operating conditions depart from the training distribution, while physics-informed neural networks typically assume fully known governing equations and dense measurements. These assumptions are rarely satisfied in production environments. To bridge this gap, we introduce the Causality-Based Artificial Neural Network (CBANN), a modular neural surrogate that embeds structural, functional, and distributional expert knowledge directly into the network architecture rather than through loss-based penalty terms. Given a directed acyclic graph describing the causal structure of the process and qualitative expert information on each dependency, CBANN represents each causal mechanism by a dedicated neural module that enforces known monotonicities, exact conservation when applicable, and distributional priors on latent variables. Although the underlying functional mechanisms are unknown, this construction is consistent with the interventional semantics of structural causal models and is designed to mitigate practical identifiability issues under partial observability. CBANN is evaluated on two synthetic benchmarks. On a simulated semi-batch reactor with sparse observations, it improves predictive accuracy over a black-box MLP while enforcing mass conservation to numerical precision when the corresponding constraint is imposed. And on a simulated nonlinear structural causal model with latent variables, architectural monotonicity and constraints stabilize latent recovery across random initializations while maintaining competitive downstream predictive performance. An out-of-distribution stress test further demonstrates improved robustness to unseen input conditions. Overall, CBANN provides an interpretable and physically consistent surrogate model that integrates expert knowledge and exhibits more reliable behavior in low-data and out-of-domain regimes.
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $3,170
OpenAlex ID: W7233540585
