<p>Modern engineering systems are increasingly instrumented with sensors for real-time monitoring and decision-making. However, sensor measurements are frequently corrupted by noise, restricting their utility for control and diagnostics. Physics-based models, while interpretable and structurally consistent, often fail to capture unmodeled dynamics and measurement disturbances, reducing their predictive fidelity in real-world deployments. Purely data-driven approaches, conversely, tend to generalize poorly when trained on limited or noisy datasets. To address these challenges, we propose a physics-guided denoising framework that integrates physics-based model predictions with a statistical energy-based model to systematically reduce signal noise. The framework is first validated on benchmark numerical examples, including the simple harmonic oscillator and the advection-diffusion equation, under varying noise levels and missing-physics conditions. Its performance is compared against existing denoising methods such as Gaussian filtering, wavelet denoising, and a physics-informed neural network trained directly on noisy data, where the proposed denoiser demonstrates superior performance in terms of root mean square error, coefficient of determination, and signal-to-noise ratio relative to ground truth. The denoiser is then applied to a synthetic laser powder bed fusion (LPBF) process model for single- and multi-track thermal history prediction, where noise is induced in high-fidelity finite element simulations as ground truth and subsequently denoised using a semi-analytical model’s guidance. The approach is further evaluated on experimentally collected thermal data from a real LPBF system. In both cases, the proposed denoiser outperforms baseline denoising methods across a range of processing conditions and enables robust, real-time interpretation of low-cost sensor data, with direct implications for predictive process control and defect mitigation in additive manufacturing.</p>