Beyond Optimization: Exploring Novelty Discovery in Autonomous Experiments (2508.20254v1)

Abstract: Autonomous experiments (AEs) are transforming how scientific research is conducted by integrating artificial intelligence with automated experimental platforms. Current AEs primarily focus on the optimization of a predefined target; while accelerating this goal, such an approach limits the discovery of unexpected or unknown physical phenomena. Here, we introduce a novel framework, INS2ANE (Integrated Novelty Score-Strategic Autonomous Non-Smooth Exploration), to enhance the discovery of novel phenomena in autonomous experimentation. Our method integrates two key components: (1) a novelty scoring system that evaluates the uniqueness of experimental results, and (2) a strategic sampling mechanism that promotes exploration of under-sampled regions even if they appear less promising by conventional criteria. We validate this approach on a pre-acquired dataset with a known ground truth comprising of image-spectral pairs. We further implement the process on autonomous scanning probe microscopy experiments. INS2ANE significantly increases the diversity of explored phenomena in comparison to conventional optimization routines, enhancing the likelihood of discovering previously unobserved phenomena. These results demonstrate the potential for AE to enhance the depth of scientific discovery; in combination with the efficiency provided by AEs, this approach promises to accelerate scientific research by simultaneously navigating complex experimental spaces to uncover new phenomena.

• The paper introduces INS2ANE, a framework that shifts autonomous experiments from pure optimization to systematic novelty discovery.
• The methodology combines real-time novelty scoring, strategic non-uniform sampling (SANE), and deep kernel learning to enhance exploration in complex experimental spaces.
• INS2ANE demonstrates reduced prediction error and increased data diversity compared to traditional optimization-driven approaches, opening avenues for unforeseen discoveries.

Novelty Discovery in Autonomous Experiments: The INS2ANE Framework

Motivation and Background

Autonomous experiments (AEs), integrating AI with automated laboratory platforms, have become central to accelerating scientific discovery in materials science and related fields. The prevailing paradigm in AEs is optimization: AI agents, typically using Bayesian optimization (BO) or active learning, are tasked with efficiently searching parameter spaces to maximize or minimize predefined physical properties. While this approach has yielded significant advances in targeted materials design and characterization, it is fundamentally limited in its capacity for open-ended discovery. Specifically, optimization-centric AEs are prone to oversampling regions near known optima, thereby neglecting underexplored or anomalous regions where novel physical phenomena may reside.

The paper introduces INS2ANE (Integrated Novelty Score-Strategic Autonomous Non-Smooth Exploration), a framework designed to shift the focus of AEs from pure optimization to systematic novelty discovery. The approach is motivated by the need to uncover unexpected or previously unknown phenomena, particularly in complex experimental spaces where the most scientifically valuable results may not align with predefined objectives.

INS2ANE: Framework and Methodology

Novelty Scoring

INS2ANE incorporates a real-time novelty scoring system that quantifies the uniqueness of each experimental result relative to the accumulated dataset. Five distinct novelty assessment methods are evaluated:

• Distance to Centroid (DtC): Measures global deviation from the dataset mean.
• Nearest Neighbors (NN): Assesses local density, prioritizing points in sparse regions.
• Isolation Forest (IF): An ensemble-based anomaly detection method sensitive to global outliers.
• One-Class SVM (OC-SVM): Learns a boundary around the majority of data, flagging points outside as novel.
• Local Outlier Factor (LOF): Detects local outliers by comparing local density.

The selection of the novelty metric is shown to be domain-dependent. For instance, DtC and OC-SVM are less effective in the presence of low-SNR data, while NN and IF provide more robust novelty detection in the context of ferroelectric domain mapping.

Strategic Sampling (SANE)

Traditional BO acquisition functions inherently bias sampling toward regions of high predicted reward or uncertainty, which can result in redundant measurements and limited exploration. INS2ANE integrates the SANE (Strategic Autonomous Non-Smooth Exploration) algorithm, which introduces a non-uniform cost function to actively promote exploration of under-sampled or poorly characterized regions, regardless of their immediate promise according to conventional criteria. This mechanism ensures that the AE does not become trapped in local optima and systematically probes the broader experimental space.

Deep Kernel Learning Integration

The framework leverages deep kernel learning (DKL) to model the relationship between structural (image) and spectroscopic (hysteresis loop) data. DKL combines neural network feature extraction with Gaussian process (GP) uncertainty quantification, enabling both accurate prediction and principled uncertainty estimation for guiding exploration.

Experimental Validation

Pre-Acquired Dataset Evaluation

The authors validate INS2ANE using a large, annotated dataset of band excitation piezoresponse spectroscopy (BEPS) measurements on PbTiO3, comprising 10,000 hysteresis loops with known ground truth. Three AE strategies are compared:

1. Scalarizer-driven AE: Optimizes a physical descriptor (e.g., loop area).
2. Novelty-driven AE: Optimizes a selected novelty score (NN or IF).
3. INS2ANE AE: Combines novelty scoring with SANE-based strategic sampling.

Key findings include:

• Novelty-driven AEs sample a broader diversity of regions, including both in-plane and out-of-plane domains, and avoid overfitting to high-signal regions.
• INS2ANE exhibits dense sampling around local maxima of novelty, followed by strategic shifts to new regions, resulting in stratified exploration.
• Quantitative metrics: Novelty-driven and INS2ANE AEs achieve lower normalized mean error (NME) in predicting physical descriptors compared to scalarizer-driven AEs, indicating improved physical understanding. Variability in measured hysteresis loops is consistently higher for novelty-driven approaches, confirming increased exploration of distinct phenomena.

Real-World Autonomous Microscopy

INS2ANE is implemented in an autonomous scanning probe microscopy platform (AEcroscopy) for real-time experimentation on ferroelectric thin films. The results mirror those from the model dataset:

• Scalarizer-driven AEs rapidly converge to known optimal regions (e.g., domain walls), neglecting other domains.
• Novelty-driven and INS2ANE AEs maintain higher variability in measured data, with INS2ANE showing episodic increases in variability corresponding to the discovery of new novel regions.

Implications and Theoretical Considerations

The INS2ANE framework demonstrates that integrating novelty-driven objectives and strategic sampling into AEs can substantially enhance the diversity of explored phenomena, leading to more comprehensive physical models and increased likelihood of discovering previously unobserved behaviors. The approach addresses a fundamental limitation of optimization-centric AEs by decoupling exploration from predefined objectives and enabling systematic investigation of the unknown.

Theoretically, the work highlights the importance of balancing exploitation (optimization) and exploration (novelty discovery) in autonomous scientific discovery. The observed complementarity between novelty scoring and strategic sampling suggests that adaptive, meta-learning strategies—potentially incorporating human-in-the-loop oversight—could further optimize the trade-off between depth and breadth of exploration.

Practical Considerations and Future Directions

• Metric Selection: The choice of novelty metric must be tailored to the specific experimental context, with attention to noise characteristics and the nature of the phenomena of interest.
• Computational Overhead: Real-time novelty scoring and strategic sampling introduce additional computational requirements, particularly for large-scale or high-throughput experiments.
• Integration with Human Expertise: The authors propose dynamic switching between novelty-driven and optimization-driven modes, potentially guided by human intuition or meta-learning controllers, as a promising avenue for future research.
• Generalizability: While demonstrated in ferroelectric materials characterization, the INS2ANE framework is broadly applicable to other domains where open-ended discovery is critical.

Conclusion

INS2ANE represents a significant methodological advance in autonomous experimentation, enabling systematic novelty discovery through the integration of real-time novelty scoring and strategic sampling. The framework is empirically validated to increase the diversity of explored phenomena and improve physical model accuracy relative to conventional optimization-driven AEs. The results underscore the necessity of moving beyond optimization in autonomous scientific discovery and point toward future systems that dynamically balance exploration and exploitation, potentially in collaboration with human experts. This paradigm has the potential to accelerate the identification of new physical phenomena and deepen scientific understanding across a range of experimental sciences.