Abstract
Deep Kernel Learning (DKL) has emerged as a powerful framework for Bayesian Optimization (BO), via combining expressive representation learning models with typical Gaussian Processes (GPs) surrogate models. However, conventional DKL typically relies on weight-based feature extractors (e.g., multilayer perceptrons (MLPs)), which often lack interpretability and may suffer from overfitting under data scarcity or training instability, potentially leading to a degraded uncertainty quantification in GP models. Grounded in the Kolmogorov-Arnold representation theorem, this paper proposes a novel DKL-KAN framework that employs Kolmogorov-Arnold Networks (KANs) as adaptive feature extractors, formulating a DKL-KAN surrogate model. Unlike MLPs, the KANs learn data-driven univariate functions, yielding more sample-efficient and stable representations for regression under limited data regimes. Followed by the GP, the DKL-KAN facilitates end-to-end learning of expressive latent representations while maintaining robust uncertainty quantification. The proposed DKL-KAN framework is further validated with the Williams-Otto (W-O) process benchmarks. Compared to classical GPs and DKL-MLPs, the DKL-KANs consistently exhibit superior predictive accuracy and uncertainty calibration even under high-dimensional and scarce dataset. The results show that the KANs generate more structured and informative latent embeddings, thus enhancing the ability to capture complex physical nonlinearities in sparsely sampled edge regions of the design space where traditional models often falter. Moreover, the DKL-KAN is embedded into the BO and yields the efficient optimization performance on the W-O process benchmark, achieving faster early-stage improvement and converging to higher objective values. These results demonstrate the potential of DKL-KAN surrogates for accelerating BO in chemical engineering processes.