Revolutionizing Solvent-Base Selection with Parametrized Optimization

81% Yield Achieved in 85% Fewer Experiments

Quick summary

Learn how Sunthetics maintains security and compliance standards for its AI driven experiment optimization platform.

Industry

Pharmaceutical and Materials R&D

Client

A synthetic chemistry research group optimizing conditions for a Suzuki–Miyaura cross-coupling reaction—a staple in pharmaceutical and materials R&D. The team faced challenges identifying ideal solvent–base pairs from a vast design space.

Challenge

• Identify the optimal solvent–base combination to maximize product yield (Ar1–Ar2) using minimal experiments.
• Navigate the complexity of categorical variables (solvents and bases), which cannot be optimized using traditional numerical approaches.
• Avoid inefficient combinatorial screening that previously required 81 experiments to find a >75% yield.

Goal

• Reach or exceed a 75% yield.
• Achieve this with fewer than 15 experiments.
• Uncover non-obvious or previously overlooked solvent–base combinations.

Approach & Solution

• Used SuntheticsML’s Bayesian Optimization, powered by proprietary Supervised Learning (SL) and Active Learning (AL).
• Encoded categorical variables with real chemical meaning:
  • Solvents described by dielectric constant and polarity.
  • Bases described by pKa and ionization energy.
• ML algorithm generated predictions and selected the most promising combinations for testing, iterating every two experiments.

Results & Metrics

• Final best-performing combination:
  • EtOH + KOtBu → 81% yield
• Performance progression:
  • Started at 7% yield with Toluene + NaOtBu
  • Gradual increase across 4 model iterations, ending at 81%
• Total number of experiments:
  • Only 12, guided across 5 iterations
• Algorithm-recommended combinations:
  • Included candidates that researchers had not previously considered or had erroneously ruled out
• Corrected experimental error:
  • An earlier false negative (0% yield for EtOH–KOtBu) was overturned thanks to the model’s suggestion, revealing experimental mislabeling
• Compared to empirical (combinatorial) approach:
  • 81 experiments needed to achieve 83% yield (DME + KOH)
  • SuntheticsML achieved 81% in just 12 experiments
  • → 85% experiment reduction

The Sunthetics Edge

“SuntheticsML not only guided to an unexpected high-yield result but also flagged an experimental error that would've led them astray. It saved time, materials, and gave us confidence in our choices.”

Key Takeaways

• Categorical optimization is solvable: Proper parameterization unlocks categorical ML use cases.
• ML sees what humans miss: EtOH–KOtBu was dismissed by researchers but identified as optimal by the model.
• Fewer iterations, better outcomes: Just 5 modeling iterations were needed to reach 81% yield.
• Data-light, insight-rich: ML-based parameter tuning beats trial-and-error even with a small dataset.
• Major reduction in experimentation: 85% fewer tests, less waste, and faster progress.