RESEARCH SPOTLIGHT
Hi, I am
Luca Calcado
HOST INSTITUTE
Aston University
PROJECT TITLE
New architectures of optical reservoir computing and extreme learning machine.
KEY WORDS
Photonic neural networks; reservoir computing; extreme learning machines; semiconductor optical amplifiers; all-optical channel equalisation
RESEARCH AIM
My research aims to build optical neural networks from standard telecom amplifiers, without delay loops, so that fibre-optic links can correct their own distortion in the optical domain instead of burning power on digital signal processing.
RESEARCH OVERVIEW
Communication networks and AI both run on electronics that burn a lot of power. My research builds computers that process information using light instead. Data already travels through optical fibre as light, so the idea is to do the computation while the signal is still light, before converting it to electricity. I design small optical circuits made from standard telecom parts, mainly optical amplifiers, that behave like a neural network and learn from examples. The first goal is repairing distorted signals in fibre links, a job currently done by power-hungry electronic chips.
RESEARCH CHALLENGE
Optical hardware has no intrinsic memory unless a delay loop is added, and no way to apply trainable weights without converting back to electronics. The challenge is to obtain both memory and trainable nonlinearity from the physics of a semiconductor optical amplifier, so that a photonic network equalises a distorted telecom signal end-to-end in the optical domain, at accuracy competitive with a digital equaliser and a fraction of the energy.
RESEARCH INNOVATION
Most optical reservoir computers get their memory from a fibre delay loop, which fixes the memory depth at fabrication and adds latency. I get memory from the carrier dynamics of the semiconductor optical amplifier itself, so the same device provides the nonlinearity, the memory, and a trainable response. The network is a Kolmogorov-Arnold architecture rather than a fixed random reservoir: the activation functions are the learnable elements, and each one maps onto a physical MZI-SOA-VOA module. This gives a hardware network with far fewer trainable parameters than an equivalent multilayer perceptron, and every parameter corresponds to a physically settable bias or attenuation rather than a number that must be approximated in optics afterwards.
RESEARCH IMPACT
I have only simulation work till today, so having chance to hands on work in a great lab makes me excited
RESEARCH SUMMARY
Optical reservoir computing and extreme learning machines process signals in the analogue domain, but conventional implementations use fibre delay loops that fix the memory depth and add latency. I develop delay-loop-free optical architectures built from semiconductor optical amplifiers (SOAs), where carrier dynamics provide both the nonlinearity and the memory. The building block is an MZI-SOA-VOA module operating directly on optical power. From it I construct photonic Kolmogorov-Arnold networks (SSP-KAN) in which the learnable activation functions are realised in hardware rather than in software. The work combines physically calibrated device models, large-scale simulation of how system hyperparameters set memory depth and accuracy, and experimental validation. The target application is all-optical channel equalisation for PAM4 and coherent links, with an optically weighted readout that removes the electronic bottleneck at the output layer.
RESEARCH VISUAL
RESEARCH JOURNEY
RESEARCH SECONDMENT
IFISC (CSIC–UIB), Spain (2 months) & Thales Research & Technology, France (1 month)
During my secondments at IFISC (CSIC–UIB), Spain, and Thales Research & Technology, France, I will strengthen both my academic and industrial expertise in neuromorphic photonics. At IFISC, I will explore the photonic implementation of reservoir computing devices and investigate algorithms designed specifically for photonic hardware rather than adapted from digital machine learning. This secondment will also provide training in learning methods for physical systems where gradients are unavailable or unreliable, benchmarking approaches within the photonic reservoir computing community, and hyperparameter analysis for memory depth in non-delay architectures. At Thales Research & Technology, I will gain insight into the industrial applications of neuromorphic photonics, learning how device-level performance is evaluated against practical engineering requirements such as packaging, stability, manufacturing yield, and system-level constraints that determine whether laboratory demonstrations are suitable for real-world deployment.
LEARNING GOALS
At IFISC, the training methods that work when the network is a physical device. In simulation I can backpropagate through a calibrated SOA model, but on hardware the model is never exact, and the group there has spent years on the algorithms that handle that gap. At Thales, the opposite perspective: what an industrial developer asks of a device before taking it seriously, which is not the question a paper asks.
RESEARCH BENEFITS
The project has three parts that I currently do in sequence: model the SOA, train in simulation, then transfer to hardware. The transfer step is where the accuracy is lost. IFISC works on exactly this problem for photonic reservoirs, so the secondment addresses the weakest link rather than adding a side topic. The Thales visit sets the target: all-optical equalisation is only interesting if it beats a digital equaliser under realistic conditions, and the industrial view of those conditions should shape which experiments are worth running in the final year.
CAREER DEVELOPMENT
They cover the two directions I could take after the PhD. IFISC is an academic group working on the theory and algorithms side of physical computing, which is the route if I stay in research; Thales is a photonics industrial lab, which is the route if I move to R&D. Doing both during the PhD means the choice is made on experience rather than on assumptions about what each environment is like. Practically, both build the same core skill: getting a physical device to do reliable computation, which is what any job in this area requires.
MORE THAN PHD
Why did you decide to join POSTDIGITAL+ ?
The project topic matched what I wanted to work on, which is rarer than it sounds — most PhD adverts in photonic computing are either device fabrication or pure machine learning, and this one is the architecture question in between. Aston has the fibre and nonlinear photonics background the SOA work needs. The network structure mattered too: the industrial mentorship and secondments are built into the contract rather than something you have to arrange yourself later.
What do you think is the biggest advantage of being part of the International Doctoral Network?
Access to hardware and methods no single group has. Photonic computing is split across platforms — integrated silicon, fibre systems, laser networks, free-space — and a PhD in one lab usually means learning one of them. In POSTDIGITAL+ I can compare my SOA-based architecture against integrated reservoirs at UGent and coupled-laser systems at CSIC, using people who built them rather than their papers. That changes what counts as a fair benchmark. The second advantage is calibration: seeing what other DCs consider a hard problem tells you quickly whether your own difficulty is fundamental or self-inflicted.
What skill have you developed the most so far?
Building simulations that stay physically honest. Early on I wrote models that trained well and described a device that could not exist — negative powers, gain without saturation, parameters outside anything a real amplifier does. Most of my code now enforces the physics as a constraint rather than checking it afterwards. The related skill is knowing when a result is too good, which is usually a sign the model has quietly stopped being a model.
What are your career ambitions after completing your doctorate?
A postdoc in photonic computing, then a group of my own working on hardware-native learning algorithms. The field is at the point where the architectures are ahead of the training methods, and that gap is where I want to work.
What inspired you to pursue this research Area?
My MSc was on an opto-electronic Ising machine, looking at how bit resolution in the electronic feedback limits what the optical part can solve. That was the first time I saw a physical system doing computation directly, and also the first time I saw the bottleneck sitting in the conversion between optics and electronics rather than in either domain alone. Removing that conversion is what my PhD is about. The earlier pull was more basic — my BSc work was laser characterisation, and I liked that in optics you can usually see the physics you are measuring.
What surprised you most since starting your PhD?
The biggest surprise was how much of the work is modelling, and how much accuracy depends on getting the device model right rather than the network architecture. I expected the split to be the other way round. The second surprise was the review process — my first paper was rejected and then accepted on appeal, which taught me more about how to make an argument than the original writing did.