High-dimensional sparse recipe formulation in process industries requires modelling nonlinear associations among many covariates, while respecting sparsity constraints and multi-criteria tolerances. The problem appears in settings where only a small number of ingredients or parameters should be active in any given recipe.

The paper introduces a physics-informed hybrid forward-inverse machine learning framework for reconstructing such recipes. The abstract positions this approach against traditional physics-based classical methods, though the description of those methods is cut off in the available text.

Because the source abstract is truncated, details on the framework's architecture, experimental results, and comparisons are not available here. The contribution, as stated, is the combination of physics-informed constraints with a forward-inverse learning strategy for a sparse, high-dimensional industrial optimisation problem.