(530a) Composite Trust Region Bayesian Optimization for High-Dimensional Agroecosystem Model Calibration and Local Sensitivity Analysis

Environmental and agroecosystem models are essential tools for understanding and predicting the impact of land management and climate on soil, water, and crop dynamics. These process-based simulators capture complex, nonlinear biogeochemical interactions, but parameter calibration remains a key bottleneck due to their high dimensionality, nonlinear structure, and computational expense. Compounding the challenge, these simulation models are often closed-source or legacy codebases, meaning that internal gradients are inaccessible (necessitating derivative-free optimization strategies). Accurate calibration is critical for producing reliable predictions, yet many models are still tuned via trial-and-error or inverse modeling routines that can be slow, fragile, or sensitive to initial guesses. Scalable and robust calibration methods are increasingly needed to support data-driven decision-making in agriculture, environmental policy, and climate resilience.

Several approaches have been developed to address this challenge. Gradient-based methods, such as finite-difference approximations, are widely used but require many model evaluations and are prone to numerical instability in high-dimensional or noisy settings [1]. The widely adopted PEST software [2] uses a Gauss-Marquardt-Levenberg algorithm with a linearized Jacobian to minimize a weighted sum-of-squares loss. While effective in many cases, PEST is restricted to differentiable, squared-error objectives and can struggle with nonconvex or ill-conditioned problems. Evolutionary algorithms and genetic methods [3, 4] offer more flexibility but are often prohibitively sample-inefficient. Bayesian optimization (BO) has emerged as a more sample-efficient alternative, particularly in expensive black-box settings [5]. Recent work, such as the BOA framework [6], applies SAASBO to agroecosystem model calibration and shows promising results – but assumes the existence of sparse, axis-aligned structure and incurs high computational overhead from hyperparameter inference via Hamiltonian Monte Carlo.

To overcome these limitations, we propose TuRBO-C, a composite trust-region Bayesian optimization algorithm for simulation-based model calibration. TuRBO-C extends the TuRBO method [7], which improves the scalability of BO by using multiple local Gaussian process (GP) surrogates constrained within trust regions. Our key innovation is to generalize TuRBO to composite objectives, where each component of the loss function corresponds to a different class of observations or sensor types (such as nutrient leaching, gas fluxes, or yield measurements), each modeled with a separate GP surrogate. These models are aggregated using a flexible composite loss function, allowing users to tailor the objective to the characteristics of each measurement (e.g., squared loss, quantile loss). The trust region is then adapted using a weighted average of the GPs’ lengthscales, effectively balancing exploration across heterogeneous outputs with different sensitivities. We also introduce a simple yet effective local sensitivity analysis method based on surrogate-based Sobol index estimation within the trust region surrounding the best-found parameter set, enabling interpretable insights into parameter relevance.

We demonstrate the effectiveness of TuRBO-C on a real-world calibration task using the DeNitrification-DeComposition (DNDC) model, a biogeochemical simulator widely applied to assess greenhouse gas emissions, nutrient cycling, and agroecosystem performance. We use field-scale data from an agricultural site in Ohio [8], incorporating time-series observations from multiple output types to calibrate DNDC’s high-dimensional parameter space. TuRBO-C outperforms traditional and modern baseline methods in both predictive accuracy and evaluation efficiency. Moreover, the composite surrogate models and sensitivity analysis provide actionable insight into how different parameter subsets influence specific outputs, highlighting the value of this approach for both model calibration and interpretability in complex environmental systems.

References:

[1] Bard, Y. (1970). Comparison of gradient methods for the solution of nonlinear parameter estimation problems. SIAM Journal on Numerical Analysis, 7(1), 157-186.

[2] Doherty, J. (1994). PEST: a unique computer program for model-independent parameter optimisation. In Water Down Under 94: Groundwater/Surface Hydrology Common Interest Papers; Preprints of Papers (pp. 551-554). Barton, ACT: Institution of Engineers, Australia.

[3] Cao, H., Recknagel, F., & Orr, P. T. (2013). Parameter optimization algorithms for evolving rule models applied to freshwater ecosystems. IEEE Transactions on Evolutionary Computation, 18(6), 793-806.

[4] Wang, Q. J. (1997). Using genetic algorithms to optimise model parameters. Environmental Modelling & Software, 12(1), 27-34.

[5] Frazier, P. I. (2018). A tutorial on Bayesian optimization. arXiv preprint arXiv:1807.02811.

[6] Scyphers, M. E., Missik, J. E., Kujawa, H., Paulson, J. A., & Bohrer, G. (2024). Bayesian Optimization for Anything (BOA): An open-source framework for accessible, user-friendly Bayesian optimization. Environmental Modelling & Software, 182, 106191.

[7] Eriksson, D., Pearce, M., Gardner, J., Turner, R. D., & Poloczek, M. (2019). Scalable global optimization via local Bayesian optimization. Advances in neural information processing systems, 32.

[8] Bhattarai, A., Steinbeck, G., Grant, B. B., Kalcic, M., King, K., Smith, W., ... & Khanal, S. (2022). Development of a calibration approach using DNDC and PEST for improving estimates of management impacts on water and nutrient dynamics in an agricultural system. Environmental Modelling & Software, 157, 105494.