Real-time Multi-instrument Autonomous Materials Screening for Discovery of Novel Phase-change Memory Materials

Technical Summary: Real-time Multi-instrument Autonomous Materials Screening for Discovery of Novel Phase-change Memory Materials

Problem Statement
Autonomous laboratories (ALs) have revolutionized materials discovery by integrating automated execution, data analysis, and decision-making. However, a significant bottleneck remains in the integration of diverse, heterogeneous data streams from multiple instruments. Current autonomous systems typically operate as single-instrument, single-property closed loops, where decision-making occurs independently across characterization equipment. Furthermore, the elucidation of synthesis-process-structure-property relationships (SPSPR) often relies on post-experiment analysis after data collection is complete, rather than during live experiments. This sequential approach fails to leverage the potential correlations between structural and functional properties in real-time, slowing down the discovery of complex materials like phase-change memory (PCM) systems, particularly those involving mixed-phase regions and non-equilibrium phases.

Methodology: The Multi-instrument Autonomous Discovery (MAD) Framework
The authors propose the Multi-instrument Autonomous Discovery (MAD) framework, a system designed to perform parallel autonomous experimentation (AE) by combining structural property mapping and functional property optimization in a closed-loop manner.

• Experimental Setup: The framework was demonstrated on a Mn-Sb-Te (MST) ternary thin-film library, a system previously unexplored for PCM applications. Two compositionally identical spreads were prepared: one annealed to a crystalline state for X-ray diffraction (XRD) and one kept in an amorphous state for electrical resistance measurements. These were mounted on separate instruments (a diffractometer and a contact probe station) connected to a central server via a client-server model.
• Data Fusion and Modeling: The core of MAD is a multi-output Gaussian Process (GP) model employing a co-regionalization kernel (specifically, an Intrinsic Coregionalization Model, ICM).
  • Structural Decomposition: XRD patterns were pre-processed and decomposed using Non-negative Matrix Factorization (NMF) into seven "end-member" components (representing potential elemental, binary, and ternary phases) and their corresponding abundances. This converts the phase identification problem from discrete labeling into a continuous regression task, capturing mixed-phase characteristics.
  • Joint Prediction: The NMF-derived phase abundances (7 outputs) and electrical resistance measurements (1 output) were concatenated into an 8-dimensional dataset. A multi-output GP with a rank-one latent structure was trained on this data to model the shared covariance between structural and functional properties.
• Decision Making: The system utilizes distinct acquisition functions for each task within the active learning loop:
  • Exploration (XRD): Maximum Uncertainty (MU) was used to maximize knowledge of the crystal structure distribution.
  • Exploitation (Resistance): Expected Improvement (EI) was used to identify the composition with the maximum electrical resistance in the amorphous state ($R_{amo}$), a key figure of merit for PCM.
  • The central controller performs joint inference to capture cross-property correlations while making independent decisions for the next experimental points for each instrument.

Key Results

• Efficiency and Speed: The MAD framework successfully identified promising electrical PCMs and mapped the SPSPR within 25 closed-loop iterations. This represents a seven-fold speed-up compared to conventional grid mapping and a five-fold speed-up compared to random sampling. The entire live experiment, including phase mapping and materials optimization, was completed in approximately 5 hours.
• Accuracy: With only ~42 measurements, the system reconstructed XRD patterns with 95% cosine similarity to experimental data. For materials optimization, the system converged to the optimal composition (maximum $R_{amo}$) in fewer than 25 data points, with the minimum regret distance converging to zero.
• Material Discovery: The study identified Mn$_{28}$Sb$_{52}$Te$_{20}$ as the composition with the highest amorphous resistance and Mn$_{14}$Sb$_{43}$Te$_{43}$ as yielding the largest resistance contrast ($\Delta R \approx 5$ orders of magnitude). The analysis revealed that the trigonal MnSb$_2$Te$_4$ phase (indexed as components C0 and C2) is central to the observed phase-change behavior.
• Benchmarking: In silico benchmarks comparing MAD against independent GPs and random sampling demonstrated that the coregionalized kernel significantly outperformed independent models in both phase mapping accuracy (measured by Dynamic Time Warping) and optimization convergence. The shared latent space allowed for coordinated learning, preventing the "idling" of simpler tasks while complex ones converged.

Significance and Claims
The paper claims that MAD establishes a new paradigm for large-scale autonomous facilities by enabling experiments to run in parallel rather than independently. By integrating multi-modal data streams through a shared latent representation, the framework allows for:

1. Real-time SPSPR Elucidation: The ability to make decisions based on shared knowledge between structure and function during the experiment, rather than waiting for post-hoc analysis.
2. Handling Mixed Phases: The use of NMF-derived continuous abundances allows the model to capture the physics of mixed-phase regions and gradual transitions, which discrete labeling approaches often miss.
3. Accelerated Discovery: The framework significantly reduces the time and resource expenditure required to explore vast compositional spaces, as demonstrated by the rapid identification of novel PCM candidates in the MST system.

The authors position MAD as a versatile and modular foundation for future co-orchestrated multi-instrument environments, capable of handling heterogeneous data streams to achieve faster convergence and better materials co-design criteria.