Tuning of Atomic Layer Deposition Pulse Time through Physics-Informed Bayesian Active Learning

This paper introduces a physics-informed Bayesian active learning framework that integrates a Langmuir adsorption model into a Gaussian Process to autonomously and efficiently tune Atomic Layer Deposition pulse times, achieving faster convergence, higher accuracy, and reduced precursor usage compared to standard data-driven approaches.

The Gist

Imagine you are trying to bake the perfect cake, but you have a very strict rule: you must stop adding ingredients the exact second the batter reaches the perfect consistency. If you stop too early, the cake is raw; if you stop too late, it's overcooked.

In the world of advanced manufacturing, specifically Atomic Layer Deposition (ALD), scientists face a similar challenge. They are building ultra-thin films (like the coating on a solar cell or a medical implant) one atom at a time. To do this, they spray a chemical "precursor" into a chamber. The goal is to spray it just long enough to cover every single atom on the surface (saturation), but not a second longer, because that wastes expensive chemicals and time.

Traditionally, finding this "perfect spray time" was like guessing in the dark. Engineers would run hundreds of experiments, wasting tons of expensive chemicals, just to find the right timing. It was slow, expensive, and inefficient.

This paper introduces a smart new way to solve this problem using a "Physics-Informed AI." Here is how it works, broken down with simple analogies:

1. The Problem: Guessing in the Dark

Imagine trying to find the exact moment a sponge is fully soaked. You could just keep pouring water and checking every second. But the water is expensive, and the sponge is tiny. If you guess wrong, you waste water and time.

• The Old Way: Run a test, check the result, guess a new time, run another test. Repeat 50 times.
• The Result: You eventually find the answer, but you've wasted a fortune in chemicals.

2. The Solution: The "Smart Guide" (Physics-Informed AI)

The authors created a computer program that acts like a smart guide who knows the laws of physics. Instead of just looking at the data (the "what"), the AI also knows the "how" and "why" based on a famous scientific rule called the Langmuir Isotherm.

Think of the Langmuir model as a map of how sponges absorb water. The AI doesn't just guess; it uses this map to predict where the "full saturation" point is likely to be.

3. The Secret Sauce: The "Two-Stage Filter"

Real-world experiments are messy. Sensors make mistakes, and timing isn't perfect. It's like trying to hear a whisper in a noisy room.

• The Innovation: The AI uses a two-step strategy:
  1. The Noise Filter (The Gaussian Process): First, the AI looks at the messy, noisy data and smooths it out, like a photo editor removing grain from a blurry picture. It creates a clean, clear line of what probably happened.
  2. The Physics Fit (The Langmuir Model): Then, it takes that clean line and fits the "physics map" onto it to extract the true numbers.

Analogy: Imagine trying to read a handwritten note that is covered in coffee stains.

• Standard AI: Tries to guess the letters based on the stains. It gets confused.
• This New AI: First, it digitally removes the coffee stains (smoothing). Then, it uses its knowledge of handwriting styles (the physics model) to read the letters perfectly.

4. The Results: Fast, Cheap, and Accurate

The researchers tested this "Smart Guide" in computer simulations and then in a real lab using Titanium Dioxide (a material used in electronics).

• Speed: The AI found the perfect spray time in 5 tries. A standard AI might have needed 20 or more.
• Efficiency: It used 2 to 4 times less chemical than the old methods.
• Accuracy: For high-quality targets (95% coverage), it was nearly perfect.
• Robustness: Even when the data was very noisy (like a stormy day), the AI still found the right answer because it trusted the "physics map" more than the messy data points.

5. The Catch (and why it's okay)

When the team tested the AI on lower saturation levels (like 80% coverage), the predictions weren't perfect.

• Why? The "physics map" (Langmuir model) assumes that once a spot is filled, it stays filled. But in the real world, some chemicals might "slip off" (desorb) if the timing isn't perfect.
• The Takeaway: The AI is excellent for the most important goal: getting the surface fully covered (95%+). For lower coverage, the real world is a bit more chaotic than the simple map, but the AI is still much better than guessing.

Summary

This paper is about teaching a computer to be a master chef instead of a random guesser. By combining real-world data with the fundamental laws of how chemicals stick to surfaces, the new system learns faster, wastes less money, and builds better materials. It's a step toward factories that can tune themselves automatically, saving time and resources for the future of technology.

Technical Summary

1. Problem Statement

Atomic Layer Deposition (ALD) is a critical technique for depositing high-quality thin films with atomic-scale precision. However, determining the optimal precursor pulse time required to achieve surface saturation is a resource-intensive process.

• Current Limitations: Traditional optimization relies on trial-and-error experimental campaigns that consume significant precursor materials and machine time.
• Data-Driven Gaps: While standard machine learning (e.g., Gaussian Processes or Neural Networks) can predict saturation, "black-box" models often require extensive data to converge, struggle with extrapolation beyond the training window, and lack physical interpretability.
• Goal: To develop an autonomous framework that minimizes precursor usage and experimental iterations while accurately identifying saturation conditions, even in the presence of experimental noise.

2. Methodology

The authors propose a Physics-Informed Bayesian Active Learning (BAL) framework that integrates domain-specific physical knowledge directly into the machine learning algorithm.

A. Physical Model: Langmuir Isotherm

The framework utilizes a modified Langmuir adsorption model to describe the growth rate $y(x)$ as a function of exposure time $x$:
$$y(x) = \frac{G_{max}Kx}{1+Kx} - D(x)$$

• $G_{max}$: Maximum surface coverage.
• $K$: Equilibrium constant.
• $D(x)$: Desorption term (minimized in simulations but addressed experimentally via extended purging).
• Rationale: In ALD, parameters like temperature and pressure are often fixed; the primary variable is pulse time. The Langmuir model captures the characteristic saturation curve of surface reactions.

B. Core Innovation: Two-Stage Parameter Estimation

A key methodological novelty is the decoupling of noise filtering from physical parameter extraction:

1. Stage 1 (Smoothing): A Gaussian Process (GP) is trained on noisy experimental data. The GP acts as a sophisticated filter, generating a dense, noise-free posterior mean prediction.
2. Stage 2 (Fitting): The Langmuir parameters ($K_{phys}, G_{phys}$) are fitted to the smoothed GP predictions rather than the raw noisy data.

• Benefit: This prevents the instability often caused by fitting non-linear physical models directly to sparse, noisy data.

C. Physics-Informed Kernel

Instead of a standard statistical kernel, the authors employ a Physics-Informed Matérn Kernel.

• The input $x$ is transformed via the Langmuir function $\phi(x)$ before calculating covariance.
• This forces the GP to learn in a space that respects the underlying physics of adsorption, guiding the active learning strategy toward the most informative regions (e.g., the steepest slope of the isotherm) rather than exploring uniformly.

D. Active Learning Workflow

The system operates iteratively:

1. Initialize: Latin Hypercube Sampling (LHS) provides initial survey points.
2. Train & Predict: Fit the GP and compute posterior uncertainty.
3. Explore: Select the next sampling point ($x_{next}$) that maximizes posterior variance (uncertainty sampling).
4. Update: Measure new data, update the GP, and re-estimate physical parameters.
5. Converge: Repeat until the saturation time for a target coverage (e.g., 95%) is predicted with high confidence.

3. Key Results

A. Simulation Studies

The framework was tested across four simulated regimes: Fast Saturation, Moderate Saturation, Slow Saturation (requiring extrapolation), and High Noise.

• Convergence Speed: The physics-informed model converged to the true saturation time within 5 iterations, whereas standard data-driven GPs required significantly more.
• Prediction Accuracy:
  • Fast Saturation: Reduced prediction error from ~66% (Pure Matérn) to 17.3% (Physics-Informed).
  • High Noise: Reduced error from ~57% to 26.1%.
  • Slow Saturation: Demonstrated superior extrapolation capabilities, maintaining accuracy even when the target lay outside the initial search window.
• Resource Efficiency: Achieved a 2x to 4x reduction in precursor usage (cumulative exposure time) compared to standard approaches.

B. Experimental Validation

The framework was validated experimentally using TiO₂ deposition via Tetrakisdimethylamido Titanium (TDMAT) and Ozone on silicon wafers.

• Setup: Used in situ spectroscopic ellipsometry for feedback.
• Performance:
  • Successfully identified the pulse time for 95% saturation (5.217 s) with only 0.6% error in Growth Per Cycle (GPC).
  • The model accurately predicted the maximum GPC (0.045 nm), matching literature values.
• Observations on Lower Saturation: Deviations increased at lower saturation targets (80–90%). The authors attribute this to non-ideal desorption effects ($D(x)$) which become more significant at lower surface coverages, a nuance the simple Langmuir model does not fully capture but which provides valuable insight into system kinetics.

4. Key Contributions

1. Novel Two-Stage Strategy: Introduced a hierarchical approach where a GP smooths data before physical parameters are extracted, significantly improving robustness against experimental noise.
2. Physics-Informed Kernel: Demonstrated that embedding the Langmuir transformation directly into the GP kernel accelerates convergence and enables reliable extrapolation beyond the training data.
3. Resource Efficiency: Proven ability to reduce precursor consumption by up to 75% and identify optimal process parameters in fewer than 5 experimental cycles.
4. Experimental Proof-of-Concept: Validated the framework on a real-world ALD system (TDMAT/O₃), confirming its practical utility for autonomous process tuning.

5. Significance

This work represents a significant step toward autonomous, closed-loop ALD process development. By bridging the gap between data-driven machine learning and physical chemistry, the framework:

• Accelerates R&D: Drastically reduces the time and cost required to develop new ALD recipes.
• Enhances Sustainability: Minimizes the waste of expensive and often toxic precursor materials.
• Generalizability: The modular, open-source nature of the code allows researchers to adapt the framework to different chemistries by swapping the kinetic model, making it a versatile tool for the semiconductor and materials science industries.

In conclusion, the paper establishes that integrating fundamental physical laws (Langmuir kinetics) into active learning algorithms yields a superior optimization strategy that is faster, more accurate, and more resource-efficient than purely data-driven approaches.