Data-science driven autonomous process optimization

Abstract

Autonomous process optimization involves the human intervention-free exploration of a range process parameters to improve responses such as product yield and selectivity. Utilizing off-the-shelf components, we develop a closed-loop system for carrying out parallel autonomous process optimization experiments in batch. Upon implementation of our system in the optimization of a stereoselective Suzuki-Miyaura coupling, we find that the definition of a set of meaningful, broad, and unbiased process parameters is the most critical aspect of successful optimization. Importantly, we discern that phosphine ligand, a categorical parameter, is vital to determination of the reaction outcome. To date, categorical parameter selection has relied on chemical intuition, potentially introducing bias into the experimental design. In seeking a systematic method for selecting a diverse set of phosphine ligands, we develop a strategy that leverages computed molecular feature clustering. The resulting optimization uncovers conditions to selectively access the desired product isomer in high yield.

Fig. 1. Phosphine ligand influence on a palladium-catalyzed stereoselective Suzuki–Miyaura coupling to generate the stereoinversion product (2-Z) or stereoretention product (2-E).

^aConditions: 10 µmol 1-E, 1 µmol 1,3,5-trimethoxybenzene, 20 µmol (3-(benzyloxy)phenyl)boronic acid 3, 0.4 µmol Pd(ACN)₂Cl₂, 30 µmol K₃PO₄ (0.5 M aq) in ACN (0.05 M), 2 h at 25 °C. ^b,cConditions: 10 µmol 1-E, 1 µmol 1,3,5-trimethoxybenzene, 11 µmol (3-(benzyloxy)phenyl)boronic acid 3, 0.2 µmol Pd(ACN)₂Cl₂, 0.4 µmol L, 30 µmol K₃PO₄ (0.5 M aq) in ACN (0.05 M), 2 h at 25 °C. Ligand structures are provided in Fig. 4. Tabulated results are provided in Supplementary Information Table SI-9.

Fig. 2. Closed-loop system for autonomous optimization in batch.

The three main components to enable this closed loop include (1) ChemOS to coordinate experiments and data-driven approaches, (2) Chemspeed SWING for automated experimental setup, and (3) Agilent 1100 to characterize the experimental outcomes.

Fig. 3. Standard experimental conditions for reproducibility testing.

Conditions: 10 µmol 1-E, 1 µmol 1,3,5-trimethoxybenzene, 15 µmol (3-(benzyloxy)phenyl)boronic acid 3, 0.3 µmol Pd(ACN)₂Cl₂, 0.4 µmol L2, 30 µmol K₃PO₄ (0.5 M aq) in ACN (0.05 M), 2 h at 20 °C. Average yields for first set of replicates: 2-E: 30(±2)%; 2-Z: 19(±1)% and second set of replicates: 2-E: 28(±2)%; 2-Z: 17(±1)%. Tabulated results are provided in Supplementary Information Table SI-13.

Fig. 4. Parameters and results of optimization with ligands selected through chemical intuition in campaign 1.

Conditions: 10 µmol 1-E, 1 µmol 1,3,5-trimethoxybenzene, 10–20 µmol (3-(benzyloxy)phenyl)boronic acid 3, 0.1–0.5 µmol Pd(ACN)₂Cl₂, 0.05–2 µmol L, 30 µmol K₃PO₄ (0.5 M aq) in ACN (0.05 M), 2 h at 10–40 °C. Tabulated results are provided in Supplementary Information Table SI-10.

Fig. 5. K-means clustering on the first four principal components of the molecular descriptor set for 365 commercial monodentate phosphines.

The chemical space is represented by a two-dimensional plot of the first two principal components. Each cluster is represented by color and highlighted boxes indicate selected ligands. Selected ligand structures are provided in Fig. 6.

Fig. 6. Parameters and results of optimization with ligands selected through descriptor clustering in campaign 2.

Conditions: 10 µmol 1-E, 1 µmol 1,3,5-trimethoxybenzene, 15 µmol (3-(benzyloxy)phenyl)boronic acid 3, 0.1–0.5 µmol Pd(ACN)₂Cl₂, 0.05–2 µmol L, 30 µmol K₃PO₄ (0.5 M aq) in ACN (0.05 M), 2 h at 10–40 °C. Tabulated results are provided in Supplementary Information Table SI-11. An animated chart of E-product yield over time is provided as Supplementary Movie 2.

Fig. 7. Three-dimensional plot of three continuous parameter selections color coded for 2-E yield in campaign 2.

PhSPhos results are outlined; all other ligand results are not outlined.

Fig. 8. Two-dimensional plots of 2-E yield for each experiment ID and average 2-E yield for each sample bias value in campaign 2.

a Green indicates positive, exploitative sample bias values, while purple indicates negative, explorative sample bias values. PhSPhos results are outlined; all other ligand results are not outlined. b Green indicates positive, exploitative sample bias values, while purple indicates negative, explorative sample bias values.

Fig. 9. The maximum yield of 2-E obtained for each monodentate ligand explored in campaigns 1 and 2 mapped onto the chemical space of 365 commercial monodentate phosphines.

The chemical space is represented by a two-dimensional plot of the first two principal components. Color indicates the maximum yield of 2-E obtained for each evaluated monodentate phosphine ligand.

Fig. 10. Manual experiments under optimized conditions with predicted E-selective ligands.

Conditions: 10 µmol 1-E, 1 µmol 1,3,5-trimethoxybenzene, 15 µmol (3-(benzyloxy)phenyl)boronic acid 3, 0.4 µmol Pd(ACN)₂Cl₂, 0.9 µmol L, 30 µmol K₃PO₄ (0.5 M aq) in ACN (0.05 M), 2 h at 35 °C. Comparisons of the predicted and experimental yields are provided in Supplementary Information Table SI-16.