Researchers at the National Institute for Materials Science (NIMS) in Japan have successfully engineered a van der Waals crystal capable of mimicking neuronal cell behavior using light-induced electrical responses. By leveraging the unique electronic properties of molybdenum disulfide (MoS2), the team demonstrated synaptic-like plasticity, offering a potential path toward hardware-level neuromorphic computing that bypasses traditional CMOS-based energy constraints.
Breaking the von Neumann Bottleneck with Light
Modern computing relies on the von Neumann architecture, which physically separates the processing unit from the memory. This “bottleneck” forces data to shuttle back and forth, consuming massive amounts of power—a major hurdle for training large-scale Transformer-based LLMs. The NIMS discovery utilizes a thin-film van der Waals crystal that acts as both a sensor and a processor.
By applying light pulses to the MoS2 crystal, the material exhibits persistent photoconductivity, a phenomenon where the internal resistance changes based on the history of light exposure. This is functionally analogous to synaptic weight adjustment in a biological brain. Unlike traditional silicon transistors, which require constant power to maintain a state, this crystal retains its “memory” through structural electronic trapping, drastically lowering the energy required for inference tasks.
Material Science Meets Neuromorphic Logic
The core of this breakthrough lies in the van der Waals heterostructures, which allow for the precise stacking of atomic layers. Because these materials are held together by weak intermolecular forces rather than covalent bonds, engineers can “design” the crystal’s properties at an atomic scale.
The NIMS team, led by researchers in the field of soft-matter electronics, utilized the material’s ability to trap charge carriers when exposed to specific wavelengths. When the light is removed, the trapped carriers result in a long-lasting electrical current. This effectively creates an “artificial neuron” that integrates temporal information—a fundamental requirement for spiking neural networks (SNNs).
| Feature | Traditional CMOS | van der Waals Crystal | |
|---|---|---|---|
| Energy Efficiency | Low (High leakage) | High (Non-volatile state) | |
| Architecture | von Neumann (Separated) | In-Memory Computing | |
| Stimulus | Electrical Gate | Photo-Electrical |
The Silicon Valley Perspective on Substrate Innovation
While the academic result is significant, the path to commercialization remains fraught with manufacturing challenges. Scaling van der Waals crystals from a laboratory-grade wafer to mass-produced silicon-compatible substrates requires solving the “transfer” problem—getting these delicate atomic layers onto standard wafers without introducing defects.
“We are seeing a shift where the substrate itself is becoming the computer. The NIMS approach is compelling because it treats light not just as data input, but as the control signal for the logic gate itself. However, the industry is still waiting for a CMOS-compatible process that doesn’t involve manual, small-scale exfoliation,” says Dr. Elena Rossi, a semiconductor process engineer and independent tech consultant.
Ecosystem Implications: Beyond the GPU
If this technology matures, it threatens to disrupt the current dominance of NVIDIA’s H100 and Blackwell architectures. Current AI hardware relies on massive parallelization of arithmetic logic units (ALUs). A van der Waals-based neuromorphic chip would operate on a fundamentally different principle: analog signal processing that mimics brain chemistry.
The development of neuromorphic software frameworks, such as Nengo or Intel’s Lava, could eventually bridge the gap between these physical crystals and high-level AI development. Should NIMS or their partners successfully integrate these crystals into a standard FPGA or ASIC workflow, the energy cost per neuron-fire could drop by several orders of magnitude.
What Happens Next?
The immediate hurdle is stability. In experimental settings, these crystals are often susceptible to environmental degradation, such as oxidation. Future research will focus on encapsulation techniques—using hexagonal boron nitride (hBN)—to protect the active layers. Until the stability matches the 10-year lifespan of enterprise-grade hardware, this remains an R&D-stage innovation rather than a threat to current data center infrastructure.
For now, the NIMS findings confirm that the future of computing may move away from pure electron flow and toward a marriage of optics and material science. As of June 2026, the focus for the research community remains on improving the sensitivity of the MoS2 crystal to lower-intensity light sources, which would enable integration with low-power mobile ARM-based SoCs.
