Bayesian optimization-driven parallel-screening of multiple parameters for the flow synthesis of biaryl compounds

The study addresses the challenge of efficiently and simultaneously optimizing multiple parameters in continuous-flow reactions, where variables such as flow rate, reactor (micromixer) type, temperature, concentration, catalyst loading, and stoichiometry interact nontrivially. Existing one-variable-at-a-time approaches are inefficient and assume independence among parameters. While Gaussian process regression and Bayesian optimization have shown promise, incorporating categorical variables (e.g., solvent, reagent, reactor type) typically requires complex descriptor engineering and modeling. The purpose here is to develop a simpler, practical Bayesian optimization (BO) strategy that can directly handle categorical parameters without feature extraction, enabling rapid identification of optimal, sustainable, low-temperature, low-catalyst-loading conditions for redox-neutral cross-coupling of iminoquinone monoacetals (IQMAs) or quinone monoacetals (QMAs) with arenols to synthesize functionalized biaryls in flow.

The authors situate their work within data-driven and automated reaction optimization in flow systems, highlighting benefits such as reproducibility, rapid mixing/heating, short reaction times, and automation. Prior work has used Gaussian process regression to predict reaction parameters from small datasets and applied BO to multi-parameter screening in batch and flow. Incorporating categorical chemical variables typically relies on molecular descriptors and quantum chemical calculations, which complicates model construction and may be impractical. Engineering features for non-numerical parameters like reactor type are also hard to quantify despite their importance. For the target chemistry, organocatalytic cross-coupling of QMAs/IQMAs with arenols has been reported using strong acids (e.g., TFA, MeSO3H) in batch with good yields and scope but drawbacks of high catalyst loading (20 mol%), long times (16 h), and high temperatures (100 °C). The present work aims to overcome these constraints via flow and BO.

- Reaction platform: Continuous-flow microreactor system evaluating redox-neutral cross-coupling of IQMAs (or QMAs) with arenols catalyzed by triflic acid (TfOH) in toluene or toluene/EtOAc. - Optimization approach: Bayesian optimization (BO) with Gaussian process surrogate and parallel lower confidence bound (LCB) acquisition function for batch-of-3 suggestions. Categorical variables are handled by one-hot encoding (e.g., micromixer type: Comet X, β-type, T-shaped), avoiding reliance on relative integer magnitudes. - Parameters screened (six total): Five continuous numerical parameters and one categorical parameter. • Stoichiometry of arenol (1–3 equiv.) • Temperature (20–60 °C; extended as needed) • Substrate concentration (IQMA or QMA; 0.01–0.1 M; adjusted for QMA) • Flow rate (0.05–0.2 mL/min) • Catalyst loading (0.5–2 mol% for IQMA; adjusted down to ~0.25–0.98 mol% for QMA) • Micromixer type (Comet X, β-type, T-shaped), with reactor volumes: Comet X 2.4 mL, β-type 2.7 mL, T-shaped 1.6 mL. - Initial dataset: Six experiments spanning the parameter space (entries 1–6 in Table 1 for IQMA; same design for QMA in Table 2), chosen to avoid expensive/toxic reagents and reduce material consumption. - BO cycles: Parallel BO proposes 3 new experiments per iteration, each suggesting a mixer and associated numeric parameters. Experimental results are fed back to retrain the model. - Objective: Maximize NMR yield of coupling product. - Experimental procedure (IQMA exemplar): Two syringes deliver solutions of 1a (0.065 mmol, 0.015 M in toluene) and 2a (3.0 equiv) with TfOH (1.5 mol%) in toluene at total flow rate 0.08 mL/min through a Comet X micromixer (2.4 mL; residence time 15 min) at 25 °C, followed by quench into sat. NaHCO3. Work-up: extraction with EtOAc, drying (Na2SO4), concentration, silica gel chromatography. For QMA optimization, solvent was toluene/EtOAc (10/1) and conditions adjusted per BO (e.g., β-type mixer, different concentrations/flow rates/temperatures and lower acid loading). - Computational resources: BO implemented using GPyOpt; scripts deposited on Zenodo (DOI: 10.5281/zenodo.7151503). - Substrate scope studies: Applied optimized IQMA conditions (Comet X, 25 °C, 1.5 mol% TfOH, 0.015 M, 0.08 mL/min) across diverse IQMAs (1b–k) and arenols (2a–j). For QMA, applied the BO-optimized β-type conditions to various arenols. - Scale-up and transformations: Gram-scale flow synthesis of 3a (1.0 g of 1a). Post-synthetic modifications: desulfonylation of 3e; Ni(II)/Zn-catalyzed reductive coupling of 5f with diphenyl phosphine oxide. - Mechanistic probe: Tested 2-methoxynaphthalene with 1a; lack of product supports necessity of mixed-acetal intermediate and [3,3]-sigmatropic rearrangement under the proposed mechanism.

- BO rapidly identified optimal flow conditions for IQMA/arenol coupling, achieving up to 96% NMR yield (entry 15, Table 1) and 93% isolated yield for 3a using Comet X micromixer, flow rate 0.08 mL/min, residence time 15 min, 25 °C, 1.5 mol% TfOH, 0.015 M concentration. - Parallel LCB acquisition was critical; other acquisition functions (single EI, LCB, parallel EI) were less effective for optimizing mixer choice. - Flow significantly outperformed batch under strong acid: batch with TfOH led to IQMA decomposition and side products, whereas rapid mixing and quenching in flow suppressed undesired pathways. - Substrate scope (IQMA): Broad, delivering 2-amino-2'-hydroxy-biaryls in good to excellent yields across electron-donating/withdrawing groups and various sulfonyl or acetal protections (typical isolated yields 68–96%). Moderate enantioselectivity was observed in a chiral phosphoric acid variant (3h: 23% ee, 78% yield). - QMA case required distinct conditions: Initial IQMA-optimized settings gave only 38% isolated yield of 5a. BO re-optimization (β-type mixer, lower TfOH loading 0.35 mol%, 30 °C, 0.044 M, 0.068 mL/min) improved yields to 69% NMR (66% isolated) for 5a (Table 2, entry 15). - QMA substrate scope: Delivered 2,2'-dihydroxy biaryls in moderate to excellent yields, including 5k at 97% and several others at 84–93% isolated yields; X-ray structure confirmed for 5j. - Gram-scale synthesis: 3a produced in 85% yield from 1.0 g of 1a under optimized flow conditions. - Downstream transformations: Removal of Ms group in 3e gave 6 in 86% yield; Ni(II)/Zn-catalyzed P-arylation of 5f afforded 7 in 60% yield. - Practical advantages: Low catalyst loading (as low as 0.35–1.5 mol% TfOH), room temperature operations, inexpensive/low-toxicity solvent (toluene), short residence times (<15–30 min), and ability to optimize categorical engineering variables (mixer type) without descriptors.

The findings demonstrate that Bayesian optimization with parallel LCB can efficiently navigate mixed continuous and categorical parameter spaces for flow reactions, enabling rapid, data-efficient identification of high-yielding conditions. Direct handling of categorical variables via one-hot encoding allowed the algorithm to propose and discriminate between micromixer types without requiring specialized physical descriptors. The optimization revealed that optimal reactor hardware is reaction-dependent: IQMA couplings favored Comet X at lower concentrations and specific residence times, whereas QMA couplings benefited from a β-type mixer and higher substrate concentrations with lower acid loading. This likely reflects differences in mixing/stirring characteristics and reaction kinetics for the two substrate classes. The flow platform mitigated issues observed in batch (acid-induced IQMA decomposition) through rapid mixing and immediate quench, leading to cleaner conversions at room temperature. The broad substrate scopes, high yields (up to 96% IQMA products and 97% QMA products), and successful scale-up underscore the method’s relevance to sustainable synthesis. Moreover, the approach reduces experimental burden versus one-variable-at-a-time strategies and showcases how BO can guide practical process decisions including reactor component selection.

A rapid, efficient, and regioselective flow synthesis of functionalized biaryls was achieved by combining organocatalytic redox-neutral cross-coupling of IQMAs or QMAs with arenols and Bayesian optimization-driven multi-parameter screening. The BO strategy directly optimized both numerical and categorical parameters (including micromixer type) without recourse to complex descriptors, enabling low catalyst loadings, room-temperature operation, short residence times, and use of inexpensive solvent. Optimal conditions translated to broad substrate scopes, high yields (up to 96% for IQMA-derived biaryls and 97% for QMA-derived biarenols), and gram-scale synthesis. This constitutes, to the authors’ knowledge, the first redox-free flow process for synthesizing highly functionalized biaryls. Future work will extend BO-assisted screening to multiple categorical parameters and pursue highly enantioselective variants using immobilized chiral catalysts in flow.

- Batch reactions under strong acid conditions led to IQMA decomposition and side-product formation, necessitating flow to achieve good outcomes. - Optimization outcomes were reaction-specific, requiring different micromixer types and concentration regimes for IQMAs versus QMAs, which may limit universal conditions. - Certain substrates (e.g., 2-methoxynaphthalene) failed under optimized conditions, consistent with mechanistic requirements for mixed-acetal formation. - Enantioselective variants showed only modest ee (e.g., 23% ee for 3h), indicating room for improvement. - Some QMA couplings provided only moderate yields despite optimization, and multiple BO iterations were needed to reach optimal performance.