SK Sunwoo Kim

AI for Chemical Engineers · Course 01

Surrogate Models for Process Decision-Making

From process prediction to optimization-compatible AI. Learn to build vector and sequence surrogates, examine physical consistency, and use trained models inside mathematical optimization.

Designed by Sunwoo Kim · Graduate entry · 8-week self-paced course
Syllabus and reading list Week 1: notes and lab

What this course asks

If a model predicts a process well, when can it also support good operating decisions? We study prediction error, physical constraints, formulation equivalence, and decision quality as separate properties that must be checked together.

Prerequisites: basic Python, calculus, linear algebra, and material balances. LP/MILP, convexity, and KKT are introduced as needed. Current materials: syllabus and Week 1. Later weeks will be released as they are completed.

Weekly materials

Week 2

Steady-state vector-to-vector surrogates

Multi-output MLPs, scaling, sampling, output errors, and balance residuals.

Planned · materials in preparation
Week 3

Dynamic sequence-to-sequence surrogates

Window MLPs, state transitions, RNN/LSTM, direct prediction, rollout, and trajectory splits.

Planned · materials in preparation
Week 4

Physics-informed learning and KKT-hPINN

Residual losses, automatic differentiation, linear equality projection, KKT derivation, and rank conditions.

Planned · materials in preparation
Week 5

Exact MILP embedding of ReLU networks

Network graphs, binary variables, Big-M derivation, valid activation bounds, and scaling.

Planned · materials in preparation
Week 6

Embedding in practice: bounds, OMLT, and horizons

Bound tightening, LP relaxations, MIP gaps, OMLT, affine projection layers, and finite-horizon ReLU models.

Planned · materials in preparation
Week 7

ICNN/PICNN and conditional LP reformulation

Convexity conditions, epigraphs, fixed context, constraint direction, and equality counterexamples.

Planned · materials in preparation
Week 8

A complete process decision workflow

Decision-point errors, extrapolation, feasibility, economic performance, simulator rechecks, and targeted data collection.

Planned · materials in preparation