Accelerating Bayesian Optimal Experimental Design via Local Radial Basis Functions: Application to Soft Material Characterization
Tianyi Chu, Jonathan B. Estrada, Spencer H. Bryngelson · 2025
Read Accelerating Bayesian Optimal Experimental Design via Local Radial Basis Functions: Application to Soft Material Characterization by Tianyi Chu, Jonathan B. Estrada, Spencer H. Bryngelson (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-13283.txt.
Introduction
Tianyi Chu, Jonathan Estrada, and Spencer Bryngelson speed up the design of experiments in this 2025 paper. Bayesian optimal design picks the test most likely to teach the most, but scoring each candidate test is pricey. Their shortcut uses local radial basis functions, smooth bumps fitted near each point.
Those bumps build faithful stand-ins for the costly model from nearby data. With a good stand-in, the expected information gain needs only a small set of samples. The team applies the method to soft material testing, where each real run is slow. The approach cuts computing while keeping the design quality high.
Catalog record
| CREATOR | Tianyi Chu, Jonathan B. Estrada, Spencer H. Bryngelson |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 67 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2505-13283
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Accelerating Bayesian Optimal Experimental Design via Local Radial Basis Functions: Application to Soft Material Characterization Tianyi Chu, Jonathan B. Estrada, Spencer H. Bryngelson Abstract We develop a computational approach that significantly improves the efficiency of Bayesian optimal experimental design (BOED) using local radial basis functions (RBFs). The presented RBF–BOED method uses the intrinsic ability of RBFs to handle scattered parameter points, a property that aligns naturally with the probabilistic sampling inherent in Bayesian methods. By constructing accurate deterministic surrogates from local neighborhood information, the method enables high-order approximations with reduced computational overhead. As a result, computing the expected information gain (EIG) requires evaluating only a small uniformly sampled subset of prior parameter values, greatly reducing the number of expensive forward-model simulations needed. For demonstration, we apply RBF–BOED to optimize a laser-induced cavitation (LIC) experimental setup, where forward simulations follow from inertial microcavitation rheometry (IMR) and characterize the viscoelastic properties of hydrogels. Two experimental design scenarios, single- and multi-constitutive-model problems, are explored. Results show that EIG estimates can be obtained at just 8 % 8\% of the full computational cost in a five-model problem within a two-dimensional design space. This advance offers a scalable path toward optimal experimental design in soft and biological materials. Keywords: Bayesian optimal experimental design; Radial basis functions; Viscoelastic material; High strain rate † † Code available at https://github.com/InertialMicrocavitationRheometry/IMR_RBF_BOED 1 Introduction Cavitation rheometry has emerged as a powerful technique for characterizing the finite deformation mechanics of soft materials and predicting their behavior under extreme conditions. Using focused energy sources, such as lasers, ultrasound, or shock waves, this method can induce inertial cavitation within compliant materials, triggering rapid and large deformations at high strain rates ( 10^{3} – 10^{8} \text{/}\mathrm{s} ). Low-strain-rate techniques ( 10^{-3} – 10^{-1} \text{/}\mathrm{s} ), including needle- and confinement-induced cavitation, enable quasi-static testing by probing material responses under gradual hydrostatic pressure or mechanical compression [ 1 , 2 ] . Together, these approaches provide complementary insights into mechanical behavior across dynamic and quasi-static regimes. These capabilities hold promise for biologic and medical applications, including tissue phantom studies, laser surgery diagnosis, and targeted cellular interventions such as DNA manipulation [ 3 , 4 , 5 , 6 , 7 ] . Characterizing realistic soft materials and biotissues, however, remains challenging under extreme loading rates. These systems often exhibit high compliance, which complicates stress-strain measurements [ 8 , 9 ] , and their nonlinear, time-dependent mechanical behavior deviates markedly from idealized linear elastic models [ 10 , 11 ] . To address these limitations, inertial microcavitation rheometry (IMR) has emerged as a robust method for characterizing soft materials at high strain rates [ 12 ] . By integrating laser-induced cavitation (LIC) experiments with physical bubble dynamics models, IMR enables accurate estimation of viscoelastic properties through time-resolved analysis of bubble radius evolution during cavitation events. This technique has been applied to characterize widely used biomimetic hydrogels, including polyacrylamide (PA) [ 12 , 13 , 14 ] , agarose [ 15 , 16 ] , collagen [ 17 ] , fibrin [ 18 ] and gelatin [ 19 ] . Recent work by Spratt et al. 2021 further enhanced IMR by integrating data assimilation techniques, providing a scalable framework for bubble-collapse rheometry. Despite these advances, a bottleneck persists in cavitation experiments: efficiency. Preparatory procedures—such as chemical treatment, degassing, and swelling protocols—required to create pristine, bubble-free samples impose significant experimental overhead [ 21 , 12 ] . The continuous variability of key material parameters, such as crosslinker density and additive concentration, compounds this challenge. To address these limitations, an optimal experimental design (OED) strategy accelerates the probing of material responses across diverse physical mechanisms, such as deformation, pressure, and thermal effects. The goal of the OED approach is to identify the most informative experiments while ensuring robustness against uncertainties—in this case, those inherent to cavitation rheometry. For this goal, we maximize the expected information gain (EIG), a Bayesian metric estimating the expected reduction in uncertainty about model parameters resulting from a prospective experiment. EIG is a relative entropy, or Kullback–Leibler (KL) divergence, between the posterior and prior distributions, measuring the information gained from observational data. By optimizing EIG, the Bayesian OED (BOED) process identifies optimal experimental configurations that maximize the informativeness of measurements while minimizing resource costs. In addition, BOED enables efficient exploration of high-dimensional parameter spaces in the presence of multiple sources of uncertainty, including variability in experimental configurations and measurement noise. Our recent work developed a sequential IMR-based BOED approach for characterizing material behavior [ 22 ] , assuming that the viscoelastic response is governed by two constitutive models: the Neo-Hookean Kelvin–Voigt (NHKV) model [ 23 ] and the generalized quadratic law KV (Gen. qKV) model [ 13 , 24 ] . When theoretical models are not available a priori to guide the design process, forward simulations across a range of candidate models become necessary. Prior studies have explored various spherical bubble dynamics models to characterize the viscoelastic properties [ 23 , 25 , 12 , 26 , 27 ] . In such scenarios, the total number of required simulations can grow substantially, making EIG computationally intensive and less likely to converge. This drawback is not an inherent limitation of OED approaches based on a choice of constitutive model. Instead, it reflects a broader challenge common to complex modeling scenarios, particularly those involving computationally expensive partial differential equations or other high-fidelity physics-based simulations. Efforts have been made to reduce the computational cost associated with BOED, particularly in high-dimensional or model-agnostic regimes. A common strategy involves using EIG estimators with lower sampling complexity. For example, the sample-reused technique reuses inner-loop Monte Carlo samples across outer loops to reduce cost from O(N^{2}) to O(N) [ 28 ] . The Monte Carlo with Laplace Approximation (MCLA) estimator [ 29 ] approximates the posterior distribution as Gaussian, greatly simplifying sampling and lowering computational costs. By integrating this into the nested Monte Carlo (NMC) framework, the Laplace-based importance sampling NMC estimator achieves higher sampling efficiency while maintaining unbias and avoiding numerical underflow [ 30 ] . Variational inference frameworks like variational OED (VOED) [ 31 ] and its normalizing flow extensions [ 32 ] amortize inference by approximating posterior or evidence distributions with parametric surrogates. These estimators enable the direct evaluation of EIG at candidate designs and are typically followed by design-space exploration using algorithms such as simulated annealing [ 33 ] , interacting particle systems [ 34 ] , stochastic optimization [ 28 , 35 , 36 , 37 ] , and Bayesian optimization (BO) [ 31 , 38 , 39 ] . Alternatively, gradient-based optimizations, such as stochastic gradient descent, can circumvent direct EIG evaluation when the EIG gradient is accessible. Despite these advances, both direct and gradient-based approaches require extensive forward model evaluations to estimate expectations, rendering them computationally prohibitive for complex scenarios that involve expensive simulations. While existing methods focus on reducing sample complexity, our work addresses a different, and often more critical, bottleneck: the cost of deterministic forward simulations. We use the radial basis functions (RBFs) to generate large sample sets from a relatively small number of forward simulations, reducing computational cost without sacrificing EIG accuracy. Rooted in scattered data interpolation, RBF-based methods offer a systematic means