91% Accuracy: Reac-Discovery Merges Math Modeling, ML, and Automation for Generalizable Labs

Background and Motivation

Self‑driving laboratory (SDL) systems accelerate catalytic reactor design but lack a unified model for geometric parameters such as porosity, surface area, and tortuosity. Traditional CFD simulations are costly and rely on expert intuition, limiting reproducibility across different chemical systems.

Reac-Discovery Platform Overview

Reac-Discovery is a semi‑autonomous digital platform that integrates mathematical modeling, machine learning (ML), and automated experimentation to address the generality problem of SDLs. The platform implements a closed‑loop workflow comprising three tightly coupled modules: Reac‑Gen, Reac‑Fab, and Reac‑Eval.

Reac‑Gen: Geometric Modeling and Parameterization

Reac‑Gen generates periodic open‑cell (POC) structures using predefined mathematical equations (e.g., Gyroid, Schwarz, Schoen‑G). Users specify three primary parameters—size (S), level threshold (L), and resolution (R)—to produce diverse topologies at macro‑ and micro‑scales. The workflow includes:

Input of key geometric parameters and construction of an implicit surface via scalar field evaluation.

Projection of the surface into 3D space, mesh generation, scaling, and cylindrical cropping to fit reactor housings.

Automatic smoothing and continuity checks to ensure printable and fluid‑compatible geometries.

Export of STL files for fabrication and XLSX files containing surface area, porosity, tortuosity, hydraulic diameter, and other descriptors.

Reac‑Fab: Printability Assessment and Fabrication

Reac‑Fab receives STL and descriptor files, predicts printability with a neural‑network classifier trained on 236 experimental samples, and conducts high‑resolution stereolithography (SLA) printing. The classifier achieves 91 % prediction accuracy, reducing material waste and experimental cost. After printing, samples undergo surface functionalization and catalyst loading.

Reac‑Eval: Automated Experimentation and Dual Optimization

Reac‑Eval integrates hardware via a unified Python interface to run parallel experiments on printed reactors. It monitors temperature, flow rates, concentrations, and NMR signals in real time. The module performs two‑stage optimization:

M1 : Machine‑learning model optimizes process variables (flow, temperature, concentration) based on experimental data.

M2 : Neural‑network model optimizes geometric descriptors, feeding the best designs back to Reac‑Gen for a second experimental round.

The overall loop enables simultaneous optimization of process parameters and reactor topology.

Dataset Generation

Reac‑Discovery autonomously creates three internal datasets:

Structural parameterization dataset (Reac‑Gen) – quantified POC designs.

Printability dataset (Reac‑Fab) – mapping of geometric descriptors to successful prints.

Performance dataset (Reac‑Eval) – real‑time NMR‑derived temperature, flow, concentration, and yield data.

All datasets are publicly available on Zenodo (https://hyper.ai/datasets/45520).

Experimental Validation

Two representative multiphase reactions were used to benchmark the platform:

Benzophenone Hydrogenation

Stage G1: Reac‑Gen generated nine Gyroid geometries; Reac‑Eval ran 60 experiments, training M1.

Stage G2: M2 incorporated geometric descriptors, selecting the optimal topology from 480 printable POCs.

Results: M1 accurately identified the optimal process region among >1 million combinations; M2 further improved prediction precision, confirming high‑accuracy dual‑optimization.

CO₂ Cycloaddition

Stage G1: 60 experimental runs produced an initial dataset; M1 predicted optimal conditions.

Stage G2: M2 jointly optimized geometry and process, selecting the best printable POC.

Results: Predicted optimal conditions matched experimental outcomes, achieving 40 %–90 % conversion across four epoxide systems, demonstrating cross‑system generalization.

Broader Impact and References

The platform’s approach aligns with recent AI‑driven self‑driving laboratory studies, such as the Nature Communications paper “Reac-Discovery: an artificial intelligence–driven platform for continuous‑flow catalytic reactor discovery and optimization” and related works on flow‑reactor design (e.g., Machine learning‑assisted discovery of flow reactor designs, Nat. Commun.; Self‑Driving Laboratories for Chemistry and Materials Science, ACS). These studies highlight the emerging paradigm where hardware automation, real‑time data analytics, and AI decision‑making form a closed‑loop scientific workflow.

Challenges noted include high system cost, limited data standardization, and model generalization, but ongoing advances in algorithmic robustness and hardware integration are expected to mitigate these issues.