Invention Reference Number
Dual-GP addresses limitations in traditional GPBO-driven autonomous experimentation by incorporating an additional surrogate observer and allowing human oversight, this technique improves optimization efficiency via data quality assessment and adaptability to unanticipated experimental challenges.
Description
The integration of artificial intelligence and machine learning (ML) with experimental automation shows promise in speeding up materials discovery and optimization. Autonomous experimentation employs AI techniques like reinforcement learning and Bayesian optimization, allowing systems to conduct experiments with limited human oversight and adapt in real-time. Gaussian Process Bayesian Optimization (GPBO) is particularly effective for navigating experimental spaces to find optimal conditions but often relies on predefined scalarizer functions, which can misguide decisions due to their inflexibility.
To overcome these limitations, we developed the Dual-GP approach, which employs a second surrogate model (e.g., Gaussian Process in this set up) to assess raw data and refine scalarizer function reliability in real-time. This new method prioritizes high-quality data, improving optimization performance and convergence while minimizing the influence of noise and unreliable data. In a synthetic dataset with varying noise, Dual-GP effectively identified noisy experimental spaces, leading to more precise outcomes compared to standard GPBO.
In cases where no reference data exists to assess quality, human oversight becomes essential. Experts can provide quality scores that train the secondary GP, enabling better adaptation to unexpected findings. This framework allows for dynamic adjustments during experimentation, ensuring that the optimization process remains focused on meaningful discoveries.
Benefits
- Real-time adjustments: Incorporates a second surrogate model that dynamically modifies the experimental space based on current data evaluations.
- Data quality assessment: Evaluates the quality of raw experimental data to ensure reliable results and minimize misinterpretations.
- Scalarizer compatibility check: Assesses how well the raw data aligns with predefined scalarizer functions, enhancing decision-making in the optimization process.
- Human-in-the-loop mechanism: Enables expert intervention when unexpected results arise, allowing for timely adjustments to the experimental setup.
Applications and Industries
- Rapid materials optimization for electronics, photovoltaics, semiconductors, etc. with quality control and human insights incorporation.
- Workflows for accelerated scientific discovery.
Contact:
To learn more about this technology, email partnerships@ornl.gov or call 865-574-1051.
