A 280-Parameter Embedding-Constrained Network for Autonomous Rate-of-Penetration Optimization on a Microcontroller

Rate of penetration (ROP) prediction is essential for drilling efficiency, yet prevailing deep learning models employ millions of parameters and rely on surface-based pipelines constrained by low-bandwidth telemetry and round-trip delays measured in minutes to hours. These delays make surface-computed optimization decisions obsolete before they reach the bit, leaving real-time drilling intelligence infeasible. We present ROP MCU_model, an encoder–regressor neural network with only 280 trainable parameters that performs autonomous ROP prediction and optimization entirely on a microcontroller. Discovered through genetic-algorithm neural architecture search and trained with supervised contrastive learning, the model imposes structural constraints on the latent embedding space to maintain accuracy under extreme parameter scarcity. Full-integer quantization compresses the network to 5.57 KB, enabling inference in 156 μs on an ESP32-S3 via TensorFlow Lite Micro. An integrated on-device optimization engine searches controllable drilling parameters within a bounded search envelope without surface communication. Validated under leave-two-well-out cross-validation on field data from the Taranaki Basin, the model achieves a mean weighted mean absolute percentage error (WMAPE) of 0.687 over nine seeded evaluation splits, the lowest prediction error among all competing models. Recomputing all normalization bounds from training-stage wells only reproduces this result within 0.2%, indicating no split-dependent leakage and a negligible effect. The on-device optimizer delivers a mean surrogate-model-predicted ROP improvement of 46.2% (median 18.5%). This work demonstrates closure of the sense–predict–optimize–decide decision loop on a commodity microcontroller; transferring the methodology to other telemetry-constrained downhole tasks remains future work.