Nicolas Smits

Experience

Research and industry roles, newest first, each with a figure from the work it produced.

At the Center for Advanced Process Decision-making I built and released SiNDAE, an open-source Python package that trains neural networks inside differential-algebraic process models with nonlinear programming solvers.

h₁, tank 1 levelh₂, tank 2 level
A₁ḣ₁ = Fᵢₙ − q₁   A₂ḣ₂ = q₁ − c₂√h₂   q₁ = NN(h₁, h₂)   The network learns the discharge q₁; the valve law of tank 2 stays a hard constraint. Peach: the three training trajectories from the paper's example; coral: the steady state.
  • Built the package on the hybrid DAE training methods developed at CAPD, so a data-driven surrogate sits inside a first-principles model and both train as one nonlinear program.
  • Wrote the documentation and worked examples, from process control to bioprocess kinetics and dynamical systems, built with Jupyter Book and published by CI.
  • Integrated further NLP solvers and tested them test-first.
  • Presented in AIChE 2026, session 542, Software Tools and Implementations for Process Systems Engineering.

Clathrate hydrates hold gas molecules such as methane and CO₂ in cages of hydrogen-bonded water. I implemented the GRADE cage-recognition algorithm in Python and improved on its functionality: it reads any .xyz frame and also recognises partly formed cages.

A 5¹² cage of 20 waters and 30 hydrogen bonds around a methane molecule, cut from the relaxed structure I methane hydrate in the project's reference files. Drag to turn it.
  • A depth-first search over the hydrogen-bond network finds five- and six-membered water rings, kept only if they pass geometric filters for planarity, convexity and interior angles.
  • Rings join into cups by adjacency and orientation, and cups pair into 5¹² and 6²5¹² cages, partly formed ones included.
  • Tuned the geometric validation across hydrate-host systems to 97% detection accuracy.
  • CH₄ and CO₂: injected into methane hydrate, CO₂ can take the methane's place in the cages, storing the CO₂ and releasing the methane as fuel. Counting cages frame by frame follows that exchange, and the tool's output fed entropy analysis for CO₂ sequestration studies.

Two summers in the global Flue Gas Treatment team at Lhoist's Business Innovation Center, where the Sorbacal® sorbents that scrub acid gases from industrial flue gas are developed.

Illustrative. The shape of the second-order model, not its numbers: conversion rises along a logistic curve as more sorbent is dosed, and as the flue gas gets hotter the curves swap, from HCl converting more to SO₂ converting more. The pilot data stays with Lhoist.
  • Built kinetic models of a pilot scrubbing reactor in Python (SciPy): ordinary differential equations coupled to nonlinear parameter estimation, fitted to the pilot runs and used to predict outcomes at new conditions.
  • Added the second-order reaction dynamics the legacy model left out, which give conversion its logistic rise with the stoichiometric ratio.
  • Extended it to secondary reactions of HCl, SO₂ and CO₂ and to temperature dependence, which captures the change from HCl to SO₂ capture as the gas gets hotter, and halved the mean squared error against pilot reactor data.
  • Tuned the stiff ODE solvers and put the Python model behind an Excel front-end (openpyxl), so engineers and experimentalists could run simulations and estimate rate constants themselves.

Teaching

  • 2026
    Systems and Toolchains for AI in EngineeringTeaching Assistant, 06-763, Carnegie Mellon

Education

  • 2027–30
    Carnegie Mellon UniversityPhD, Chemical Engineering, advisor Carl D. LairdOptimisationProcess Systems EngineeringMachine LearningStochastic ProgrammingReinforcement Learning
  • 2026
    Carnegie Mellon UniversityMS, AI Engineering, Chemical Engineering
  • 2025
    McGill UniversityBEng Chemical Engineering, minor in Applied AI