Beating High-Throughput with ML: Optimizing Categorical Variables in Cross-Coupling Chemistry

This study was conducted by a joint team from Ghent University and Sunthetics. The goal: optimize a Suzuki–Miyaura cross-coupling reaction using real-world, industrial-grade datasets—where categorical variables like solvent and catalyst dominate the design space​.

• Identify the best combination of catalyst, solvent, and base for a complex cross-coupling reaction.
  
  
• Achieve this using minimal experimentation.
  
  
• Tackle the high-dimensional complexity of categorical variables, which are often not easily optimized by standard ML or DoE tools.
  
  
• Avoid the 768 experiments typically required by High-Throughput Experimentation (HTE) to explore the full design space.
  
  

• Identify the best combination of catalyst, solvent, and base for a complex cross-coupling reaction.
  
  
• Achieve this using minimal experimentation.
  
  
• Tackle the high-dimensional complexity of categorical variables, which are often not easily optimized by standard ML or DoE tools.
  
  
• Avoid the 768 experiments typically required by High-Throughput Experimentation (HTE) to explore the full design space.
  
  

• SuntheticsML was deployed using proprietary:
  
  
  • Supervised Learning (SL) and
    
    
  • Active Learning (AL) algorithms
    
    
• No variable encoding or parameterization was required—SuntheticsML directly optimized categorical inputs like catalyst identity.
  
  
• Each optimization was initiated with 36 randomly selected experiments.
  
  
• The platform iteratively recommended 6 experiments per cycle: 5 likely to yield maximum performance + 1 random for exploration.
  
  
• Five independent "seeds" were used to evaluate repeatability and robustness.
  
  

• HTE benchmark:
  
  
  • 768 experiments required to explore full combinatorial space and find optimal conditions.
    
    
• SuntheticsML outcomes:
  
  
  • Average performance:
    
    
    • 6 iterations, 72 experiments, 91% experiment reduction
      
      
  • Best-case scenario:
    
    
    • 2 iterations, 48 experiments, 94% reduction
      
      
  • Worst-case scenario:
    
    
    • 9 iterations, 84 experiments, 89% reduction
      
      
• Insight generated:
  
  
  • Variable importance analysis showed the base had the strongest effect on yield—surprising the research team and shifting future optimization focus away from catalysts and solvents​
    .

"SuntheticsML matched the output of a fully combinatorial HTE campaign in a fraction of the time and cost—while revealing insights HTE didn't, like the true variable driving reaction performance."