Exploiting Device Nonuniformity: β-Ga₂O₃ Arrays for Integrated Sensing-Computing Bayesian Neural Networks

This work introduces a novel integrated sensing-computing approach that leverages the inherent nonuniformity in $\boldsymbol {\beta }$ -Ga2O3 photodetector arrays to implement Bayesian neural networks (BNNs). While conventional semiconductor manufacturing strives to minimize device variation, we demonstrate that nonuniformity can be advantageously utilized for probabilistic computing paradigms. Our fabricated $8\times 8$ array of $\boldsymbol {\beta }$ -Ga2O3 photodetectors exhibits a photocurrent variation with a standard deviation of $0.235~\boldsymbol {\mu }$ A around a mean value of $1.97~\boldsymbol {\mu }$ A, corresponding to a coefficient of variation of 11.9%, while maintaining uniform low dark current (7111 pA). These devices show excellent photoresponse characteristics with a responsivity of 5.7 A/W, an external quantum efficiency (EQE) of 2.8% $\boldsymbol {\times }10^{{3}}$ %, and a detectivity of $2.4~\boldsymbol {\times }10^{{13}}$ Jones under 254-nm illumination. The sensing-computing integrated architecture intelligently samples these nonuniform devices to form inference layers for Bayesian computation, eliminating the need for complex random number generators. The resulting BNN demonstrates superior performance compared to conventional fully connected networks in Fashion-MNIST classification tasks, particularly in uncertainty quantification with accuracy exceeding 96% after ten epochs at confidence thresholds above 0.85, while standard networks achieve only 89.5%. This pioneering approach of utilizing inherent device variation as a computational asset paves the way for future ultraefficient edge computing systems that embrace rather than mitigate semiconductor nonuniformity.