Process flowsheet optimization integrating mixed-integer programming and surrogate models with adaptive uncertainty

Ibrahim Mert Kahraman

doi:10.1016/j.ces.2026.124993

Process flowsheet optimization integrating mixed-integer programming and surrogate models with adaptive uncertainty

2026-08-24

0

PRE

AI

I

Ibrahim Mert Kahraman

C

Carine Menezes Rebello

I

Idelfonso B.R. Nogueira

DOI：10.1016/j.ces.2026.124993

Original

Original request for help

Save

Abstract

En 中文

The integration of machine learning into process engineering has enabled the development of cost-effective and sustainable designs for complex chemical processes. However, such integration introduces reliability concerns, as ANN-based surrogates inherently introduce prediction uncertainty that can lead to constraint violations under plant-model mismatch. This study develops a change-point-detection-oriented adaptive uncertainty management approach to address this issue within superstructure optimization. Unlike prior approaches that employ separate surrogate models for individual unit operations or unconstrained ANN training without structural enforcement, this study represents the entire superstructure as a single surrogate model trained via a constrained NLP formulation with conditional skip connections. In this context, conditional skip connections are introduced into the network architecture to ensure that outputs corresponding to inactive equipment and flowsheet paths are exactly zero. The resulting surrogate is embedded within a Mixed-Integer Nonlinear Programming (MINLP) formulation that simultaneously optimizes process decision variables and a simplified economic objective, while accounting for surrogate prediction uncertainty through disjunctive programming. The proposed uncertainty quantification approach introduces adaptive backoffs that reflect varying prediction accuracy across different operating regimes, reducing the likelihood of product purity constraint violations and supporting not overly conservative solutions. Results showed that prediction errors in the superstructure surrogate for cumene production decreased compared to a traditional ANN-based approach. These findings demonstrated the applicability of the proposed methodology for robust large-scale process design under ANN-based surrogate model uncertainty.

Process flowsheet optimization integrating mixed-integer programming and surrogate models with adaptive uncertainty

Abstract

Journal

Organization

Cited Papers