Multi-objective Bayesian Optimization with Human-in-the-Loop for Flexible Neuromorphic Electronics Fabrication

This paper presents a human-in-the-loop multi-objective Bayesian optimization framework that successfully identifies optimal photonic curing conditions for fabricating flexible aluminum oxide-based neuromorphic electronics, effectively navigating complex parameter spaces and high failure rates to balance large capacitance-frequency dispersion with low leakage current.

The Gist

The Big Picture: Cooking with a Flashlight

Imagine you are trying to bake the perfect cake. But instead of an oven, you have to use a super-bright, high-speed camera flash to cook it in a split second. This is essentially what the researchers did, but instead of a cake, they were making flexible electronic circuits (like the ones you might wear on your skin) using a special metal-oxide "dough."

The goal was to create a specific type of electronic component called a capacitor that can act like a human brain cell (a "neuromorphic" device). To do this, they needed to "flash-bake" a thin layer of aluminum oxide.

The Problem: The "Goldilocks" Zone is Tiny

The problem with using a camera flash to bake electronics is that there are five different knobs you can turn to control the light:

1. How bright the flash is.
2. How many times it flashes.
3. How long each flash lasts.
4. How many tiny sub-flashes are inside one big flash.
5. The timing rhythm of the flashes.

If you turn these knobs just slightly wrong, the result is a disaster:

• Too weak: The "cake" doesn't bake at all (it's unconverted).
• Too strong: The "cake" burns to a crisp or melts the plastic plate underneath.
• Just right: You get a working electronic device.

Because there are over 4 million possible combinations of these five knobs, trying to find the perfect setting by guessing and checking (a "grid search") would take a human scientist their entire lifetime.

The Solution: A Smart GPS (Bayesian Optimization)

To solve this, the researchers used a machine learning technique called Bayesian Optimization. Think of this as a smart GPS for the laboratory.

• Instead of driving every single road to find the destination, the GPS learns from your previous turns.
• It says, "Okay, turning left at 2 PM got us stuck in traffic, but turning right at 2:05 PM got us closer. Let's try a road near there next."
• The computer builds a map (a model) of what works and what doesn't, then suggests the next best experiment to run.

The Twist: The "Human-in-the-Loop" (The Taste-Tester)

Here is where the paper gets really clever. In a normal computer simulation, if a recipe fails, the computer just ignores it and moves on. But in a real lab, a "failed" experiment (a burnt film) still tells you something important: "Don't go this way, it's on fire!"

However, a computer can't look at a burnt film and say, "Oh, that's only slightly burnt, we might be able to fix it," or "That is completely uncooked." It usually just sees "Error."

The researchers added a Human-in-the-Loop (HITL).

• The Analogy: Imagine the computer is a robot chef trying to bake a cake. The robot doesn't know what "burnt" looks like. A human taste-tester (the scientist) steps in.
• The human looks at the result and gives it a score:
  • -1: Completely raw/uncooked.
  • -0.5: Undercooked.
  • 0: Perfectly baked.
  • +0.5: Slightly burnt.
  • +1: Completely charred.
• The robot learns from these scores. It realizes, "Ah, when I set the flash to this level, the human says it's 'slightly burnt.' I should avoid that area, but maybe I can get close to it."

Why this matters: Without the human, the robot kept suggesting settings that resulted in burnt films, wasting time and materials. With the human guiding it, the robot learned to avoid the "burnt" zones much faster, finding the perfect settings in fewer tries.

The Result: Finding the Perfect Balance

The researchers had two goals that fought against each other:

1. Maximize "Memory": They wanted the device to hold a lot of electrical charge that fades over time (like a human memory).
2. Minimize "Leakage": They wanted the device to not leak electricity like a broken bucket.

Usually, making a material better at holding charge makes it worse at stopping leaks. It's like trying to make a sponge that holds a lot of water but doesn't drip.

Using their "Smart GPS + Human Taste-Tester" system, they found a Pareto Frontier.

• The Analogy: Imagine a map of a mountain range. The "Pareto Frontier" is the ridge line. If you go higher up the ridge, you get a better view of the "Memory" valley, but you get closer to the "Leakage" cliff.
• The system found the perfect ridge line where they could choose exactly how much memory vs. how much leakage they wanted, depending on what the specific application needed.

The Takeaway

This paper isn't just about making better electronics; it's about how we do science.

• Old Way: Try everything, ignore the failures, and hope you get lucky.
• New Way: Use AI to guide the experiment, but let a human scientist use their intuition to explain why an experiment failed. This combination saves time, money, and materials, and helps us discover new technologies faster.

In short: They taught a computer how to bake electronics, but they let a human chef taste the burnt cookies to teach the computer what not to do next.

Technical Summary

1. Problem Statement

The fabrication of flexible neuromorphic electronics using solution-processed metal oxides (specifically aluminum oxide, Al₂O₃) on polymer substrates faces significant challenges:

• Processing Constraints: Traditional thermal annealing requires high temperatures that damage flexible polymer substrates (e.g., PET). Photonic curing (using intense xenon flash lamps) offers a millisecond-scale alternative that anneals the film while keeping the substrate cool.
• Complex Optimization Space: Photonic curing involves five interconnected input parameters (radiant energy, pulse count, pulse length, micropulse count, and duty cycle). The search space contains over 4 million combinations, making traditional grid-search methods infeasible.
• Competing Objectives: The goal is to optimize two conflicting electrical properties for neuromorphic applications:
  1. Maximize Capacitance-Frequency (C-f) Dispersion: A metric for short-term synaptic plasticity (STP), which relies on defect states.
  2. Minimize Leakage Current: Essential for transistor functionality, which requires a dense, stoichiometric dielectric with fewer defects.
• High Failure Rates: Real-world experiments often yield "failed" outcomes (e.g., unconverted films, partial conversion, or burned substrates). These failures produce no electrical data, causing standard Machine Learning (ML) models to waste resources exploring unfeasible regions of the parameter space.

2. Methodology

The authors developed a Human-in-the-Loop (HITL) Multi-Objective Bayesian Optimization (MOBO) framework to navigate this complex landscape.

• Multi-Objective Bayesian Optimization (MOBO):

  • Surrogate Models: Two Gaussian Process Regression (GPR) models were trained to predict C-f dispersion ($C_{100Hz}/C_{1MHz}$) and leakage current ($|\log(I_{leakage})|$).
  • Acquisition Functions: The study utilized the qEHVI (parallel expected hypervolume improvement) for initial rounds and a custom Pareto-UCB (Upper Confidence Bound) strategy for later rounds to balance exploration and exploitation.
  • Goal: To identify the Pareto frontier, representing the set of optimal conditions where improving one objective degrades the other.

• Human-in-the-Loop (HITL) Integration:

  • The Challenge: Standard MOBO cannot utilize "failed" experiments because they lack numerical output values.
  • The Solution: Researchers assigned a numerical score to the physical state of the film after photonic curing based on visual inspection:
    • Unconverted: -1.0
    • Partially converted: -0.5
    • Converted (Functional): 0
    • Partially burned: +0.5
    • Burned: +1.0
  • Constraint Model: A separate GPR model was trained on these scores to predict the probability of successful film conversion ($P_{constraint}$) across the input space.
  • Constrained Optimization: The acquisition function for the electrical objectives was weighted by $P_{constraint}$ ($acq_{constrained} = acq_{UCB} \cdot P_{constraint}$). This effectively "shields" the optimizer from regions likely to produce failures, directing it toward feasible processing windows.

• Interpretability: SHapley Additive exPlanations (SHAP) analysis was applied to the final GPR models to quantify the contribution of each input parameter to the objectives.

3. Key Contributions

• Novel HITL Framework: Introduced a method to incorporate subjective, qualitative human observations of experimental failures into quantitative ML workflows. Unlike binary classification (Pass/Fail), this approach captures the severity and mode of failure (e.g., under-conversion vs. burning).
• Accelerated Optimization: Demonstrated that incorporating HITL significantly reduces the number of experimental rounds required to converge on a Pareto frontier by filtering out unfeasible regions early.
• Tunable Neuromorphic Dielectrics: Successfully fabricated flexible Al₂O₃ dielectrics with tunable properties, enabling a trade-off between high C-f dispersion (for memory/plasticity) and low leakage (for transistor operation).
• Physical Insight: Used SHAP analysis to reveal the conflicting physical mechanisms driven by specific photonic curing parameters (e.g., pulse length).

4. Results

• Optimization Efficiency:
  • In initial rounds without HITL, the failure rate was high (only 5 out of 15 conditions yielded functional devices).
  • With HITL integration, the device yield increased dramatically to 9 out of 10 conditions in subsequent rounds.
  • The framework successfully identified the Pareto frontier within experimental uncertainty after five active learning rounds.
• Pareto Frontier Performance:
  • The final model identified a range of conditions where $|\log(I_{leakage})| \geq 4$ (required for functional transistors).
  • Extreme Conditions: Two distinct optimal conditions were identified:
    • Condition #40: Lowest leakage ($|\log(I_{leakage})| \approx 6.05$) but lower C-f dispersion (1.10).
    • Condition #66: Highest C-f dispersion (1.89) but higher leakage ($|\log(I_{leakage})| \approx 4.26$).
• Parameter Insights (SHAP Analysis):
  • C-f Dispersion: Favored by lower radiant energy, lower pulse count, and longer pulse length.
  • Leakage Current: Favored by higher radiant energy, higher pulse count, and shorter pulse length.
  • Key Conflict: Pulse length was identified as the most critical conflicting parameter; longer pulses increased dispersion but also increased leakage.

5. Significance

This work provides a robust blueprint for materials discovery and device fabrication where experimental failure rates are high and input parameters are highly interconnected.

• Resource Efficiency: By preventing the exploration of "dead zones" in the parameter space, the HITL-MOBO framework saves significant time, materials, and equipment usage.
• Generalizability: The methodology is applicable to any experimental workflow where results exist on a continuous spectrum of success/failure (common in materials science) and where human domain expertise can quickly identify failure modes that are difficult to automate.
• Application Impact: It enables the rapid development of flexible, wearable neuromorphic hardware, bridging the gap between complex material processing requirements and the need for efficient edge computing devices.