Abstract
Autonomous discovery in experimental science often struggles with non-smooth, noisy landscapes where standard Bayesian Optimization (BO) may converge prematurely on single, potentially spurious optima. We present SANE (Strategic Autonomous Non-Smooth Exploration), a framework designed to discover multiple local and global optima in complex parameter spaces. By integrating a cost-driven probabilistic acquisition function with a human-intervened "dynamic gate," SANE navigates experimental landscapes effectively, avoiding noise-induced traps and ensuring a comprehensive mapping of physical phenomena.
Introduction
In the pursuit of material discovery, autonomous laboratories increasingly utilize Bayesian Optimization (BO) to navigate parameter spaces. While standard BO excels at finding a single global optimum in smooth, well-behaved environments, real-world experimental data is rarely so cooperative. Two major hurdles often impede traditional BO:
- Non-Smoothness and Noise: Experimental measurements, such as spectroscopy, are frequently noisy. This noise can create "fake" optima—spikes in data that appear to be discoveries but are actually measurement artifacts.
- The "Winner-Take-All" Trap: Standard BO is designed to converge on a single peak. In scientific exploration, however, researchers often need to identify all interesting regions (local optima) to fully understand the underlying physics. A "vanilla" BO might exhaustively refine one peak while completely missing a secondary, more scientifically significant region.
The SANE Framework
SANE modifies the standard BO workflow to prioritize discovery over simple optimization through two primary mechanisms: Strategic Probabilistic Navigation and a Human-Intervened Dynamic Constrained Gate.
Strategic Probabilistic Navigation
Instead of assuming uniform exploration costs, SANE utilizes a set of metrics ($f_1, f_2, f_3$) based on the distance from previously explored locations and the quality of the output. It employs a probabilistic check to decide whether to continue exploiting a current region or "hop" to a potentially new region of interest. This strategic "hopping" mimics a researcher’s intuition, balancing the refinement of known results with the curiosity to explore distant, unknown areas.
Human-Intervened Dynamic Constrained Gate
To handle experimental noise, SANE incorporates a "human-in-the-loop" approach. At the start of an experiment, a human expert labels a small set of initial samples as "feasible" or "infeasible" (e.g., distinguishing real physical hysteresis loops from electronic noise). SANE uses these labels to build a "gate"—a surrogate model that penalizes the acquisition function in regions likely to produce junk data. This gate updates autonomously as new data is acquired, requiring minimal ongoing human intervention.
Experimental Validation and Results
The framework was validated across three distinct scenarios: a synthetic 2D Branin function, a Sm-doped $BiFeO_3$ (Sm-BFO) combinatorial library, and $PbTiO_3$ (PTO) thin film hyperspectral data.
- Multi-Optima Discovery: In synthetic tests, SANE successfully identified and exploited all three optima (one global and two local), whereas standard BO focused almost exclusively on the global peak.
- Noise Resilience: In the Sm-BFO experiment, standard BO was frequently misled by noisy, low-voltage readings in infeasible regions. Guided by the human-defined gate, SANE ignored these artifacts and focused on physically relevant regions.
- Comprehensive Coverage: In the PTO thin film study, SANE discovered a secondary optimal region with large loop areas that standard BO missed entirely. Quantitatively, SANE achieved a higher Structural Similarity Index (SSIM) to the ground truth experimental space compared to standard BO, capturing the "big picture" more accurately.
Discussion and Future Work
SANE represents a significant shift from optimization to mapping. By allowing humans to define "bad data" early on, the framework effectively blinds the AI to noisy traps, saving significant experimental resources.
However, challenges remain. The "penalty factor" ($P$) used to enforce the gate is problem-dependent and requires careful tuning. Furthermore, while SANE improves exploration strategy, future work is needed to develop predictive models specifically for non-differentiable functions and to scale these strategic "hops" to higher-dimensional parameter spaces.
Summary updated