PL-KKT-hPINN: Enforcing Nonlinear Equality Constraints on Neural Networks via Piecewise-Linear Projection

A new neural-network framework enforces nonlinear physical equations as hard constraints rather than soft penalties, eliminating post-training violations, according to a paper submitted in 2026 [1]. The method, called piecewise-linear Karush–Kuhn–Tucker hard-constrained physics-informed neural network (PL-KKT-hPINN), uses a piecewise-linear projection to guarantee that a model’s outputs satisfy nonlinear equality constraints at inference [1][2]. Standard physics-informed neural networks, or PINNs, embed governing equations only as loss-function terms during training, which means they can still violate those equations when making predictions [2]. The new framework extends an earlier KKT-hPINN approach that enforced linear equalities by orthogonally projecting outputs onto the feasible region defined by the Karush–Kuhn–Tucker conditions [2]. Researchers tested PL-KKT-hPINN on a continuous stirred-tank reactor case study with one and two inputs [1][2]. The model preserved predictive accuracy comparable to an unconstrained neural network while cutting constraint violations substantially [2]. In low-data settings, the hard-constrained network delivered lower root-mean-square error than its unconstrained counterpart for limited training sample sizes [2]. The authors describe the result as a computationally efficient and physically consistent surrogate-modeling tool for nonlinear chemical-engineering systems [2]. Chemical process models often rely on surrogate approximations to avoid expensive simulations, but soft enforcement of conservation laws can produce physically implausible results when data are scarce. The PL-KKT-hPINN design addresses that gap by embedding the constraints directly into the network’s architecture, so every forward pass respects the underlying physics [2]. The piecewise-linear projection strategy handles nonlinearities that the original KKT-hPINN could not, broadening the class of systems to which hard constraints can be applied [2]. Broader efforts to embed domain knowledge into machine-learning models have accelerated across engineering disciplines, driven by the need for reliable predictions in safety-critical settings. The United Nations Sustainable Development Goals, adopted in 2015, call for innovation in industrial infrastructure and responsible production, areas where physically consistent surrogate models could accelerate process design while reducing waste [6]. Although the PL-KKT-hPINN paper does not reference those goals, the alignment with SDG 9 — industry, innovation and infrastructure — illustrates the practical stakes of making neural-network models obey physical law [6].