Milan, Italy · October 2025A team of researchers from Politecnico di Milano, École Polytechnique Fédérale de Lausanne (EPFL), Stanford University, the University of Cambridge, and the Max Planck Institute has published a landmark study in Nature (Vol. 645, 2025) demonstrating a new approach to training physical neural networks (PNNs) that could fundamentally reshape the energy landscape of artificial intelligence.The paper, titled "Training of physical neural networks," addresses one of the most significant bottlenecks in neuromorphic and photonic computing: how to train physical neural networks effectively without relying entirely on external digital computers to perform the computationally intensive backpropagation calculations.## The Core Breakthrough: In-Situ Photonic TrainingThe researchers developed an "in-situ" training technique that allows photonic neural networks to be trained directly using light signals within the analog photonic hardware itself. By eliminating the need to digitize information during the training process, the system dramatically reduces both energy consumption and latency.The photonic chips developed at Politecnico di Milano use silicon microchips measuring just a few square millimetres. These chips utilise light interference — the wave nature of photons — to perform the complex mathematical operations, such as additions and multiplications, that are necessary for neural network computations. When light passes through carefully engineered optical circuits, the interference patterns naturally encode the results of these mathematical operations.> [!STAT] The photonic chips achieved over 90% accuracy in data classification tasks, matching traditional digital models in both accuracy and robustness — while performing computations using light rather than electricity.## Why This Matters: The Energy Problem in AIThe timing of this research is significant. As AI models have grown larger and more capable, their energy consumption has become a pressing concern. Training a single large language model can consume megawatt-hours of electricity, and the aggregate energy footprint of AI data centres is growing rapidly. The International Energy Agency has flagged AI data centre electricity consumption as one of the fastest-growing sources of energy demand globally.Traditional AI hardware relies on digital electronic processors — GPUs, TPUs, and custom accelerators — that move electrons through silicon circuits. Each operation generates heat, and the aggregate cooling requirements of large AI data centres are substantial. Photonic computing offers a fundamentally different paradigm: photons travel at the speed of light, generate minimal heat, and can carry multiple streams of information simultaneously through different wavelengths (a property called multiplexing).> [!INSIGHT] The key innovation here is not merely using light to perform inference — that has been demonstrated before. The breakthrough is training the network in situ, on the photonic chip itself, without needing to offload the computationally expensive training process to a digital computer. This closes the loop between computation and learning, which has been the missing piece for practical photonic AI hardware.## How It Works: Light as Both Computation and Learning SignalIn a traditional neural network, training involves a process called backpropagation: the network makes a prediction, the error is calculated, and that error is propagated backward through the network to adjust the weights. In digital systems, this requires storing intermediate values and performing billions of multiply-accumulate operations.In the photonic system developed by the research team, the training process works differently:1. Forward pass: Light signals encoding input data pass through the photonic chip, where optical components (waveguides, phase shifters, and interferometers) perform the mathematical operations.2. Error calculation: The output is compared to the desired result, producing an error signal.3. In-situ adjustment: Rather than sending this error back to a digital computer for processing, the error signal is fed back into the photonic chip as light, and the optical components are physically adjusted to correct the error.This approach means the entire training loop — forward pass, error calculation, and weight update — occurs within the photonic hardware, with no need to convert between optical and digital signals at each step.## A Collaborative International EffortThe research brought together leading institutions across Europe and North America:| Institution | Role ||---|---|| Politecnico di Milano | Photonic chip design and fabrication (Photonic Devices Lab) || EPFL | Photonic circuit architecture and testing || Stanford University | Neural network theory and algorithms || University of Cambridge | Materials and optical characterization || Max Planck Institute | Theoretical physics and modelling |The work was led at Politecnico di Milano by Prof. Francesco Morichetti and Prof. Andrea Melloni of the Photonic Devices Lab, building on the group's long-standing research into programmable photonic processors and integrated photonic circuits.## Toward Sustainable AI HardwareThe broader significance of this research lies in its potential to make AI hardware more sustainable. By moving computation from energy-intensive digital data centres to localised, real-time photonic processors, the approach could enable:- Edge AI processing: Real-time AI inference on devices (phones, sensors, autonomous vehicles) without cloud connectivity- Dramatically reduced energy consumption: Photonic chips consume orders of magnitude less energy than electronic processors for equivalent computations- Lower latency: Light-speed computation eliminates the delays inherent in electronic signal processing- Reduced data centre cooling demands: Photonic chips generate minimal heat compared to electronic alternatives> [!NOTICE] While the demonstrated photonic chips match digital models in classification accuracy, the current prototypes handle relatively simple tasks. Scaling photonic neural networks to the complexity of modern large language models remains a significant engineering challenge. The research represents a proof of concept for in-situ training, not a ready replacement for GPU-based AI infrastructure.## The Road AheadThis work represents a continuation of the group's long-standing research into programmable photonic processors and integrated photonic circuits. The demonstration that physical neural networks can be trained in situ — without external digital computers — removes one of the key obstacles that has prevented photonic computing from becoming a practical alternative to electronic AI hardware.As the energy demands of AI continue to grow, approaches like photonic in-situ training may become not just advantageous but necessary. The research community will need to address challenges in scaling, programmability, and integration with existing digital infrastructure before photonic AI hardware can move from laboratory demonstration to commercial deployment.## Sources- Morichetti, F., Melloni, A., et al. "Training of physical neural networks." Nature, Vol. 645, 2025.- Politecnico di Milano, Photonic Devices Lab, research highlights, polimi.it- EngTechnica, coverage of photonic neural network training research, 2025- NIH/National Library of Medicine, abstract indexing of Nature Vol. 645 study
FIRAT Editorial Board
Institutional Research Desk · Foresight Institute of Research and Translation
The collective editorial and research translation board of FIRAT, synthesising peer-reviewed evidence, policy briefs, and division milestones across our seven foundational research pillars.
