Shape-Constrained Bayesian Active Learning of Self-Limiting Saturation Curves

Technical Summary: Shape-Constrained Bayesian Active Learning of Self-Limiting Saturation Curves

Problem Statement
Self-limiting saturation curves—monotone responses rising from zero to a plateau—govern critical processes in physical and life sciences, including gas adsorption, enzyme kinetics, dose–response pharmacology, and the growth per cycle (GPC) in atomic layer deposition (ALD). A shared bottleneck in these fields is mapping these curves from a limited number of costly measurements. Standard surrogate models, typically stationary-kernel Gaussian processes (GPs), lack structural shape constraints. When applied to sparse, noisy data, these unconstrained GPs frequently produce unphysical dips or negative slopes. These artifacts corrupt both the inferred curve and the predictive uncertainty used to select subsequent experiments, potentially leading to the selection of unphysical minima and wasted experimental resources. While post-hoc repair methods (e.g., isotonic projection) exist, they often introduce step artifacts or fail to constrain the predictive uncertainty distribution itself.

Methodology
The authors propose a model-free, Bayesian active-learning platform built on monotonic I-spline regression. The core innovation is a surrogate that enforces shape constraints by construction rather than through post-hoc correction.

• Basis Construction: The saturation response $f(x)$ is expanded as a weighted sum of Integrated Splines (I-splines), $f(x) = \sum w_m I_m(x)$. These I-splines are derived by integrating localized, non-negative M-splines (rate blocks).
• Structural Guarantees:
  • Monotonicity: By placing a Half-Normal prior on the weights $w_m$ (enforcing $w_m \ge 0$), the slope $f'(x)$ becomes a sum of non-negative terms, guaranteeing that every posterior draw is globally non-decreasing.
  • Boundary Condition: Since every I-spline vanishes at the origin ($I_m(0)=0$), the condition $f(0)=0$ holds exactly for every draw.
  • Plateau Determination: The self-limiting plateau is not imposed by a prior (e.g., a flatness constraint) but is established by the experimental design, which always includes a measurement at the maximum exposure ($x_{max}$). The plateau value is thus data-anchored.
• Active Learning Loop: The platform employs standard uncertainty sampling (selecting the next input where the predictive standard deviation is maximized) to drive the experiment selection. Because the surrogate is monotone by construction, the uncertainty band is physically meaningful, and the "level crossing" method used to read the saturation pulse time ($x_{sat}$) yields a unique, well-defined solution for every posterior draw.
• Inference: Due to the non-negativity constraints, the posterior has no closed-form solution. The authors use the No-U-Turn Sampler (NUTS) to draw hundreds of candidate curves from the posterior distribution, summarizing the fit via the mean and credible intervals.

Key Contributions

1. Monotone-by-Construction Surrogate: The paper introduces a Bayesian framework where every posterior draw is a valid, physically admissible saturation curve. This eliminates the need for post-hoc repairs and ensures the predictive uncertainty consumed by the acquisition function is physically consistent.
2. Model-Free Kinetic Adaptability: Unlike previous ALD-specific approaches that embed specific kinetic forms (e.g., Langmuir) into the kernel, this I-spline basis is agnostic to the underlying kinetics. It adapts to concave, sigmoidal, logarithmic, and dispersive shapes without re-tuning.
3. Data-Efficient Active Learning: The platform demonstrates that a fixed surrogate can recover diverse saturation curves to within measurement noise using a small budget (20 measurements), significantly outperforming random sampling.
4. Operational Read-out: A model-free method for extracting saturation pulse times via level crossing is presented, which avoids biasing the result toward a specific kinetic model.

Results
The method was benchmarked on five kinetically distinct isotherm families: Langmuir, Dissociative (Michaelis–Menten), Sips (sigmoidal), Elovich (logarithmic with a kink), and Dubinin–Radushkevich KWW (dispersive).

• Recovery Accuracy: Across all five regimes, the surrogate recovered the curves to within the detector noise floor ($\sigma_y = 0.05$). The mean saturation error across regimes was 0.029. Even for the most challenging shapes (e.g., the sigmoidal Sips onset and the kinked Elovich curve), the fit remained accurate, with the largest deviations staying below ~1.6$\sigma_y$.
• Active vs. Random Sampling: Uncertainty sampling consistently outperformed random sampling. Active learning reached noise-floor accuracy in as few as 7 measurements (for the easiest regime) and within 16 for the hardest (Sips). In contrast, random sampling only succeeded in two of the five regimes within the 20-measurement budget.
• Noise-Free Regime Sweeps: When noise was removed, the I-spline basis was found to be near-exact for all tested regimes, with saturation errors $\le 0.022$. The only exception was a step-like Sips onset ($n=4$), where the fixed log-spaced knot grid lacked resolution at the steep rise, highlighting a structural limitation of the fixed basis rather than the monotonicity machinery.
• Pulse Time Prediction: Predicted pulse times were highly accurate (within 0.02 s) up to 50% saturation. At 95% saturation, predictions were uniformly conservative (late by 1.3–1.6 s). This bias arises because the inferred plateau slightly overshoots the true value on noisy data; operationally, this is a conservative error (longer pulses deepen saturation).
• Structural Integrity: In over one million posterior draws across all experiments, not a single non-monotone draw occurred. The method successfully avoided the "dips" and "steps" characteristic of unconstrained GPs or post-hoc repairs.

Significance and Claims
The authors claim that the primary value of this work is the elimination of a failure mode rather than the invention of a new acquisition function. By embedding monotonicity into the prior, the platform ensures that the uncertainty driving the active learning loop and the resulting saturation read-out are physically trustworthy.

The paper asserts that this approach is broadly applicable wherever a self-limiting response must be learned from scarce data, spanning adsorption isotherms, enzyme kinetics, and ALD process development. The authors are modest regarding limitations, noting that the fixed log-spaced knot grid has resolution limits for extremely steep, late onsets and that the current implementation assumes single-precursor, isothermal conditions. However, they emphasize that the monotonicity machinery itself is robust and could extend to multi-parametric learning (e.g., saturation surfaces) via tensor products. The work establishes that a single, fixed, model-free surrogate can effectively replace multiple kinetic-specific models, providing a unified, data-efficient framework for characterizing self-limiting processes.