Beyond Silicon: Materials, Mechanisms, and Methods for Physical Neural Computing

A research team including Stefan Fischer, Sebastian Otte, and five other authors has published a landmark survey on arXiv, unifying the fragmented field of physical neural computing. The paper, 'Beyond Silicon: Materials, Mechanisms, and Methods for Physical Neural Computing,' argues that as silicon-based, GPU-centered AI hits energy and data-movement bottlenecks, alternative substrates like memristive devices, photonic circuits, mechanical metamaterials, and even living neural tissue offer a complementary path. These systems perform neural inference and adaptation directly in matter by exploiting intrinsic physical processes such as wave interference, elastic deformation, and biochemical regulation.

To enable cross-domain comparison, the authors introduce a first-order benchmarking scheme based on standardized tasks and physically interpretable performance dimensions like scalability, precision, and programmability. Their analysis reveals no single substrate dominates; instead, different materials occupy complementary operating regimes. For example, photonic circuits enable ultrafast signal processing, while microfluidic networks or chemical systems could enable in-sample biochemical decision-making. This framework is particularly relevant for the trend toward pervasive intelligence—deploying AI on vast numbers of resource-constrained edge devices where co-locating computation with sensors and memory is essential for efficiency.