Kinetics studies for Multi-task Bayesian Optimization
Created on December 9|Last edited on January 2
Comment
Reaction Network
I wanted to develop several reaction networks, so I could manually adjust the similarity between reactions. The idea is to manually tweak the reaction kinetics to control how similar two reactions are. I create three different cases to compare.
The hypothetical catalytic reaction network was originally developed by Connor Taylor and modified by me. In addition to the main reaction, the network includes catalyst induction to form the active species, two side reactions of starting material, a product consumption reaction and catalyst deactivation. These are all motivated by real examples seen commonly in reaction optimization. For example, homocoupling and protodeboronation are common in Suzuki couplings, and product cleavage has been reported in several industrial case studies.
Cat -> Pre-cat + Lig (k0) - InductionSM1 + SM2 + Lig -> Product + Lig (k1) - main reactionSM2 + SM2 + Lig -> Impurity1 + Lig (k2) - homocouplingSM2 -> Impurity2 (k3) - side reaction (e.g., protodeboronation)Product -> Impurity3 (k4) - cleavage of productCat + Cat -> DeactiveCat (k4)
There are up to eight ligands and five solvents to choose from for this reaction. Additionally, the temperature, residence time and catalyst concentration can be varied. The objective is to maximize a noisy measurement of yield at the end of the reaction.
The cases vary in the optimal ligand and solvent.
Case 1 optimal
In this case, L1 and S3 are the optimal combination. You can see that there is some catalyst deactivation.
L1, S3, 85°C, 120 min, 5.0 cat conc.Yield: 93.22%Conversion SM1: 82.69%Conversion SM2: 99.39%Ligand yield: 98.14%
Case 2 optimal
L8 and S1 are the optimal in this case. Also, you can see there is a significant amount of catalyst deactivation resulting in lower yield of the active ligand.
L8 S1 85°C, 60 min, 2.5 cat. concYield: 90.62%Conversion SM1: 82.96%Conversion SM2: 99.62%Ligand yield: 30.28%
Running the reaction for longer than 60 minutes or at high temperatures means product cleavage kicks in (formation of Impurity3).
L8 S1 85°C, 240 min, 5.0 cat concYield: 64.32%Conversion SM1: 83.39%Conversion SM2: 100.00%Ligand yield: 76.58%
Results
The plot below shows that MTBO does not really perform better than STBO. I wanted to investigate why.
Run set
60
Add a comment