Autonomous Reliability Qualification of Ga$_2$O$_3$-based Hydrogen and Temperature Sensors via Safe Active Learning

This paper presents a Safe Active Learning framework that autonomously characterizes the reliability of Ga$_2$O$_3$-based hydrogen and temperature sensors under coupled thermal and hydrogen stress by dynamically balancing safety constraints with experimental exploration to map device degradation and enable long-horizon forecasting.

The Gist

Imagine you have a very delicate, high-tech sensor made of a special material called Gallium Oxide ($\text{Ga}_2\text{O}_3$). This sensor is designed to detect heat and hydrogen gas, but it's fragile. If you push it too hard with too much heat or too much gas, it might break permanently.

Traditionally, scientists test these sensors by running a long, pre-planned list of experiments: "Try 300°C, then 310°C, then 320°C..." The problem is, this is slow, wasteful, and dangerous. If the sensor breaks at step 50, you've wasted 49 steps and lost the sensor.

This paper introduces a smarter way to test these sensors using a robot brain called Safe Active Learning (SAL). Here is how it works, explained through simple analogies:

1. The "Safety Guard" (The Rectification Ratio)

Think of the sensor's health like a traffic light.

• Green Light (High Rectification): The sensor is working perfectly, blocking current in one direction and letting it flow in the other.
• Red Light (Low Rectification): The sensor is damaged or degrading. It's leaking current it shouldn't.

The robot's main job is to keep the sensor in the "Green" zone. It uses a mathematical model (a Gaussian Process, which is like a super-smart weather map) to predict where the "Green" zone is and where the "Red" zone is.

2. The "Two-Phase Exploration"

The robot doesn't just guess randomly. It plays a two-round game:

• Phase 1: The Cautious Explorer.
  Imagine a hiker exploring a foggy mountain. The hiker only steps where they are 99% sure the ground is solid (safe). The robot starts by testing the sensor in mild conditions. It learns the map of the "safe" area. If the robot predicts a spot might be dangerous, it simply doesn't go there. It builds a "Trust Region"—a safe circle around the places it has already proven are safe.
• Phase 2: The Controlled Descent.
  Once the robot knows the safe boundaries, it starts to gently push the sensor toward its limits. It slowly lowers the "safety bar." It's like a trainer slowly increasing the weight on a lifter. The robot intentionally tests conditions that are almost too harsh to see exactly when and how the sensor starts to degrade. This teaches the robot how the sensor fails over time.

3. The "Time Uncertainty" Problem

In a normal computer simulation, you know exactly how long a test takes. In the real world, it's different.

• The Analogy: Imagine ordering a pizza. You know it takes about 30 minutes, but sometimes traffic makes it 45, and sometimes it's 25.
• The Solution: The robot doesn't just plan for "30 minutes." It plans for a window of time (e.g., 25 to 45 minutes). It asks: "If I start this test now, will the sensor be safe at any point during that entire window?" This prevents the robot from accidentally starting a dangerous test right before the sensor is about to overheat.

4. The "Robot Lab"

The researchers built an automated lab station (a robot arm with a probe) that does the actual testing.

• The robot changes the temperature and gas levels.
• It waits for the sensor to calm down (equilibrium).
• It runs a quick electrical test.
• It calculates the "Traffic Light" score.
• It decides where to test next, all without a human touching a button.

5. The "Crystal Ball" (Offline Forecasting)

After the robot finishes its campaign, it has a massive, high-quality dataset of how the sensor behaves. The researchers then used this data to build a long-term prediction model.

• The Analogy: Think of it like watching a plant grow for a few weeks and then using that data to predict how tall it will be in a year.
• The model they built (using a specific mathematical shape called KWW) is really good at predicting the "slow fade" of the sensor's performance. It captures the fact that sensors degrade quickly at first and then slow down, rather than just breaking suddenly.

The Bottom Line

The paper claims that this Safe Active Learning system successfully:

1. Kept the sensor safe: It only broke the sensor once (due to a weird glitch, not the algorithm's fault) during the first phase.
2. Learned the map: It figured out exactly how heat and hydrogen affect the sensor much faster than a human could.
3. Predicted the future: It used the data it collected to accurately predict how the sensor would degrade over a long period, even for conditions it hadn't tested yet.

In short, they taught a robot to be a cautious, curious scientist that learns how to break things safely so we can understand them better.

Technical Summary

1. Problem Statement

The paper addresses the challenge of characterizing the reliability of $\beta$-Ga$_2$O$_3$-based rectifying devices under coupled thermal and hydrogen stress.

• Context: $\beta$-Ga$_2$O$_3$ is a promising wide-bandgap material for power electronics, but its long-term stability is threatened by degradation mechanisms (e.g., barrier degradation, contact modification) under high temperatures and hydrogen exposure.
• Challenge: Traditional reliability testing involves executing a pre-determined matrix of stress conditions. This is inefficient for multidimensional, time-dependent operating spaces. Furthermore, standard Active Learning (AL) or Bayesian Optimization (BO) strategies prioritize uncertainty reduction, which can inadvertently drive devices into destructive operating regimes (catastrophic failure) before the model learns the safety boundaries.
• Specific Difficulty: Experimental durations are time-uncertain; the time required for a device to stabilize after changing temperature or gas concentration is unknown a priori and varies by condition. Standard BO assumes fixed evaluation times, making it ill-suited for asynchronous, long-duration experiments.

2. Methodology: Safe Active Learning (SAL)

The authors propose a Safe Active Learning (SAL) framework designed to autonomously explore the device's operating space while strictly enforcing safety constraints.

Core Components:

1. Safety Observable (Rectification Ratio):

   • Instead of optimizing performance, the algorithm uses the rectification ratio ($R$) as a proxy for device health.
   • $R$ is calculated via an intra-band comparison of forward and reverse currents around a target voltage ($V_0$).
   • A minimum threshold ($h$) is defined; falling below this indicates irreversible degradation or unsafe operation.

2. Surrogate Modeling (Gaussian Processes):

   • The rectification surface $R(t, T, G)$ is modeled using a Gaussian Process (GP) in log-space ($\log R$).
   • The kernel is an additive combination of a Squared-Exponential (RBF) term with Automatic Relevance Determination (ARD) and a linear term to capture global trends.

3. Handling Time Uncertainty:

   • Adaptive Completion-Time Window: Since experiment duration is unknown, SAL maintains a history of observed durations to construct a probabilistic window for when the next measurement will complete.
   • Time-Window Safety: Safety checks are not performed at a single nominal time but over the entire completion-time window. The algorithm ensures that the Lower Confidence Bound (LCB) of the rectification ratio remains above the safety threshold for at least 95% of the plausible completion times.

4. Two-Phase Sampling Strategy:

   • Phase 1 (Conservative Exploration): The algorithm explores the region where $R \ge h$. It uses a trust region anchored to previously verified safe conditions to prevent aggressive extrapolation. The acquisition function balances uncertainty reduction, diversity (exploring new $T, G$), and periodic revisits to track drift.
   • Phase 2 (Controlled Relaxation): As the device naturally degrades, the safety threshold is progressively relaxed (exponentially decayed) from $h$ down to $\approx 1$ (resistor-like behavior). This allows the system to map the degradation trajectory intentionally without risking catastrophic failure in the early stages.

5. Rescue Mechanism: If the safe set becomes empty (due to model pessimism or actual degradation), a rescue routine re-measures the most recent safe condition to classify the situation (modeling artifact vs. boundary behavior vs. failure).

3. Key Contributions

• Novel SAL Algorithm: Introduction of a Safe BO variant specifically tailored for time-varying, asynchronous experiments with uncertain durations.
• Experimental Validation: Successful deployment on an automated high-temperature probe station using a Pt/Cr$_2$O$_3$:Mg/$\beta$-Ga$_2$O$_3$ device. The system autonomously generated a curated, time-resolved IV dataset.
• Offline Long-Horizon Forecasting: Development of a structured GP model for post-experiment analysis. This model uses a Kohlrausch–Williams–Watts (KWW) mean function (stretched exponential) to capture saturating degradation trends, combined with a residual GP kernel for flexibility.
• Safety-First Autonomy: Demonstrated that autonomous experimentation can reduce manual burden while preserving device integrity, only intentionally pushing into risky regimes once the degradation trajectory is understood.

4. Results

• Simulation: In simulated environments, SAL successfully expanded the explored region while maintaining strict safety compliance. The GP surrogate accurately reconstructed the rectification surface, even in sparsely sampled regions, and handled added measurement noise robustly.
• Experimental Campaign:
  • Phase 1: The algorithm operated conservatively, incurring only one unsafe measurement (caused by spurious IV sweeps, not algorithmic failure). No device conditions were banned due to safety violations.
  • Phase 2: The algorithm intentionally probed lower-rectification regimes as the device degraded, successfully mapping the transition from rectifying to resistive behavior.
  • Data Quality: The campaign produced a high-quality, time-resolved dataset suitable for offline modeling.
• Offline Modeling: The KWW-based GP model, trained on the first ~133 hours of SAL data, successfully predicted device current behavior over long horizons (extrapolation) on an independent validation dataset. It accurately captured the saturating degradation trends and the systematic ordering of responses by hydrogen concentration, with uncertainty bands widening appropriately as the prediction horizon extended.

5. Significance

• Paradigm Shift: Moves semiconductor reliability characterization from static, pre-defined stress matrices to adaptive, closed-loop experimentation.
• Efficiency: Drastically reduces the time and resources required to characterize device degradation by focusing measurements on informative regions and avoiding redundant or destructive tests.
• Safety Assurance: Provides a rigorous framework for autonomous systems to operate in high-risk environments (high temp, reactive gases) without human intervention, ensuring that "safety" is mathematically guaranteed via probabilistic bounds.
• Generalizability: While demonstrated on Ga$_2$O$_3$, the SAL framework is applicable to any device class where a measurable, physics-motivated safety observable can be defined (e.g., batteries, other sensors, or materials under stress).

In conclusion, this work establishes a robust pipeline for safe, autonomous reliability qualification, proving that machine learning-driven experimentation can not only accelerate data collection but also generate the high-fidelity datasets necessary for accurate long-term degradation forecasting.