Loughborough built a memristor chip that predicts chaos, reconstructs data, reads digits, and runs logic without architectural changes — using niobium oxide thin films and claiming up to 2,000x energy savings.
image from Gemini Imagen 4
Researchers at Loughborough University demonstrated a memristor-based reservoir computing chip using nanoporous niobium oxide thin films that executes four distinct AI tasks—chaotic system prediction, missing data reconstruction, MNIST digit classification, and XOR logic—without architectural reconfiguration. The chip leverages inherent structural randomness in the nanoporous film as the computational 'reservoir,' requiring training only at the output layer, which enables significant energy savings compared to software-based approaches. While the team claims up to 2,000x lower energy consumption, this figure originates from their press release rather than direct experimental comparison in the paper itself.
A chip that runs four different AI tasks without being rebuilt sounds like a press release. This one has a paper.
Researchers at Loughborough University built a memristor-based reservoir computing chip using nanoporous niobium oxide thin films and put it through its paces: predicting the behavior of a chaotic system, reconstructing missing data, identifying pixelated handwritten digits, and running XOR logic operations. One device, no architectural changes between tasks. The results appeared in Advanced Intelligent Systems on February 18, 2026.
The efficiency claim is the one that travels: up to 2,000 times lower energy consumption compared to software-based reservoir computing on time series prediction tasks. The number comes from Loughborough's press release, which cites the paper. The paper itself compares against established benchmarks but does not reproduce that multiplier directly. Call it the research team's characterization of their result rather than an independently verified metric.
Reservoir computing is a stripped-down approach to neural networks. You throw your input at a physical system, in this case a memristor array with inherent structural randomness from the nanoporous film, and let the device's native dynamics do the computation. The reservoir is the hardware itself. You train only the output layer, which is cheap. The appeal is that the physics does the heavy lifting, so you skip the energy cost of software-based neural networks doing the same work.
The team, led by Dr. Pavel Borisov, a Senior Lecturer in Physics at Loughborough, used the Lorenz-63 system — a standard test case for chaotic dynamics — to evaluate short-term prediction accuracy. They also tested data reconstruction and digit recognition on the MNIST handwritten digit dataset, a standard machine learning benchmark. XOR logic, a fundamental Boolean operation that demonstrates general-purpose computation capability, completed the set.
The broader context is that reservoir computing has been a niche within a niche — mostly software simulations on standard hardware, with a few physical implementations that required specific conditions to work. A memristor-based version that operates at room temperature in principle using a well-understood oxide material is a meaningful step toward something you could actually build a product around. Niobium oxide is not exotic. Thin-film deposition is a standard fab process. That is the part that matters for scaling.
As Loughborough put it in its press release, the work shows that physical reservoir computing devices can handle multiple task types without architectural reconfiguration — a departure from the conventional approach of dedicating hardware to a single function.
The Engineering and Physical Sciences Research Council funded the work. It is open access at doi.org/10.1002/aisy.202500833.
Story entered the newsroom
Assigned to reporter
Research completed — 4 sources registered. Loughborough nanoporous niobium oxide memristor acts as physical reservoir computing substrate. Four tasks demonstrated from one device: Lorenz-63 cha
Draft (447 words)
Reporter revised draft based on fact-check feedback
Approved for publication
Published (423 words)
Get the best frontier systems analysis delivered weekly. No spam, no fluff.
Space & Aerospace · 3h 9m ago · 4 min read
Space & Aerospace · 5h 4m ago · 3 min read
