The geometric configuration of micro-calorimetric flow sensors strongly influences their sensitivity and linear dynamic range. Traditional design workflows often rely on static heuristics or exhaustive Finite Element Method (FEM) simulations. These approaches are computationally expensive and frequently lead to suboptimal sensor placement. To overcome this limitation, we propose an autonomous Physics-Informed Machine Learning (PIML) framework for geometric optimization. The framework integrates a reduced-order digital twin—solving the one-dimensional convection-diffusion equation—with a Bayesian Optimization (BO) engine based on Gaussian Process regression. Using an Expected Improvement acquisition function, the system efficiently identifies the globally optimal heater-sensor separation distance with minimal evaluations. Results show that the optimal sensor location is dynamic, shifting downstream as flow velocity increases to follow the elongating thermal wake. Moreover, optimizing for the absolute signal magnitude provides a superior signal-to-noise ratio compared to traditional derivative-based sensitivity metrics, especially in saturation regimes. The framework was validated against 3D numerical simulations and experimental data with less than 5% error. It also identified a critical design point at a microchannel height of 200 μm, enabling rapid, fluid-specific sensor synthesis.