CDI Project: State Space Discovery and Data-Driven Modeling of Chemical Systems

State Space Discovery and Data-Driven Modeling of Chemical Systems

Return to Data and Integration Activities

PI: Carlos M. Ortiz Marrero

Project Team: Michael S. Hughes, Francesco Luzi, Jenna Pope

Project Term: May 2018 to September 2020

Key Science Questions:

  • What data-driven tools describe time-evolving chemical systems?
  • How can data-driven models take advantage of chemical knowledge and problem structure?
  • What tools should be applied when dealing with limited experimental data?

Project Description: This project has developed a variety of data-driven models that extract insights about chemical processes. The project goal is to identify, extend, and implement data-driven tools that leverage knowledge of the underlying chemical system of interest for a spectrum of experimental scenarios (e.g., limited data, big data). The project team has:

Recurrent neural network architecture of iodine model (left); prediction example of the normalized concentration curves (right).
Recurrent neural network architecture of iodine model (left); prediction example of the normalized concentration curves (right).
  • utilized delay-embedding theorems (e.g., Takens’ Theorem) to get an appropriate reconstruction of the chemical system with limited data
  • developed predictive data-driven models for large dimensional iodine systems using a chemistry-informed neural network
  • developed a classification algorithm to detect annealed 3D printed components using sparse ultrasonic measurements obtained from limited samples

Project Publications: 

Typical specimen used to acquire data for our classification study (top); structural image constructed using the acquired ultrasonic data (bottom). The team showed it is only necessary to acquire the data along the midline (red line) to classify the various components.
Typical specimen used to acquire data for our classification study (top); structural image constructed using the acquired ultrasonic data (bottom). The team showed it is only necessary to acquire the data along the midline (red line) to classify the various components.
  • Bilbrey J, C Ortiz-Marrero, M Sassi, A Ritzmann, N Henson, M Schram. 2020. “Tracking the Chemical Evolution of Iodine Species Using Recurrent Neural Networks.” ACS Omega, vol. 5, no. 9, pp. 4588–4594, 10.1021/acsomega.9b04104
  • Luzi F, M Fenn, J Christ, Z Kennedy, T Varga, M Hughes, C Ortiz-Marrero. 2020. “Acoustic Properties of 3D Printed Polyetherketoneketone Components,” Journal of the Acoustical Society of America (Submitted March 2020)
The project team utilizes Takens’ Theorem to reconstruct an approximate trajectory from the Brusselator system (right) only using information from one coordinate (left).
The project team utilizes Takens’ Theorem to reconstruct an approximate trajectory from the Brusselator system (right) only using information from one coordinate (left).