Level 1 — Absolute Beginner

Computers today use tiny pieces of electricity to think. The electricity moves through small chips. But this uses a lot of power.

Scientists at a school in the United States have a new idea. They want to use light instead of electricity. Light is fast and uses very little energy.

The team made a special tiny particle that is part light and part matter. The new particle can be turned on and off, just like a switch.

This work could help big computer brains called AI. AI needs a lot of power right now. With light, AI may use much less power in the future.

computer

A machine that does work with numbers and words.

electricity

Energy that moves through wires and chips.

chip

A tiny piece in a computer that thinks.

light

What makes us able to see; it travels very fast.

energy

The power needed to make things work.

switch

Something that turns power on or off.

AI

Smart computers that learn and answer questions.

future

Time that has not happened yet.

Level 2 — Elementary

A team of physicists at the University of Pennsylvania has built a new type of particle that is half light and half matter. These particles are called exciton polaritons. The team showed that they can be used as switches inside a small chip, opening the door to computers that run on light instead of electricity.

The work was led by Professor Bo Zhen and was reported by ScienceDaily on May 18, 2026. The team trapped light inside a very small space called a nanoscale cavity. There, the light could interact with an ultra thin sheet of crystal. When light and matter mixed in just the right way, the new exciton polaritons formed.

The big surprise was how little energy it took to flip the switch. The Penn group measured the energy needed at about four quadrillionths of a joule. That is far less than the power used by a small flash light bulb for a fraction of a second.

Today's AI systems use huge amounts of electricity. Data centres need water for cooling and large power lines from the grid. A computer that runs on light could shrink those energy and cooling needs by a lot. Such photonic chips could also process pictures and sound much faster, because they would not need to keep changing signals between light and electricity.

physicist

A scientist who studies the basic laws of nature.

particle

A very small unit of matter.

cavity

A small hollow space, here used to trap light.

nanoscale

Of a size around one billionth of a metre.

crystal

A solid material with atoms in a regular pattern.

joule

A standard unit used to measure energy.

data centre

A building filled with computer servers.

signal

Information that is sent from one place to another.

Level 3 — Intermediate

Physicists at the University of Pennsylvania, working in the group of associate professor Bo Zhen, have demonstrated that hybrid light matter quasiparticles known as exciton polaritons can interact strongly enough with one another to perform digital switching operations, the kind of nonlinear behaviour that underpins all programmable computation. The result, amplified by ScienceDaily on May 18, 2026, points toward a class of photonic chips that could shoulder some of the workloads now handled by graphics processing units and dedicated AI accelerators.

An exciton polariton is formed when light is confined inside a high finesse optical cavity and coupled strongly to an excitonic transition in an atomically thin semiconductor, in this case a transition metal dichalcogenide monolayer. The resulting quasiparticle inherits the speed and propagation of light from its photonic component and the interaction strength and mass of an electron hole bound pair from its excitonic component. Penn's contribution is to engineer that mixing so precisely that two polaritons in the same mode can effectively see one another and shift each other's resonance, allowing a control beam to switch a probe beam between on and off states using only the light fields involved.

The energy cost of each switching event was measured at roughly four femtojoules, or four quadrillionths of a joule, several orders of magnitude below the energy budget of comparable electronic transistors operating at the same speed. That margin matters: the largest current generation AI training runs consume tens of megawatts continuously, and the marginal cost of pushing the next generation higher is increasingly bottlenecked not by silicon scaling but by data centre power, transmission line capacity, and waste heat.

A working all photonic compute fabric remains years away. The Penn demonstration is a single switching primitive, not a fully programmable architecture, and the cryogenic cooling required for high quality polariton condensates is not yet compatible with mass deployment. The authors argue, however, that their result puts polariton based logic on a quantitatively credible footing for the first time and that the same architecture could later support analog inference on photonic neural networks and even certain kinds of on chip quantum simulation.

quasiparticle

A collective excitation that behaves like a single particle.

finesse

A measure of how well an optical cavity stores light.

monolayer

A film of material exactly one atom or molecule thick.

resonance

The frequency at which a system naturally oscillates with maximum amplitude.

transistor

A small electronic switch at the heart of every digital computer chip.

femtojoule

A unit of energy equal to one quadrillionth of a joule.

primitive

Level 4 — Advanced

Researchers in the group of Bo Zhen at the University of Pennsylvania's Department of Physics and Astronomy have demonstrated that exciton polaritons, hybrid light matter quasiparticles formed by the strong coupling of cavity photons to neutral electronic excitations in atomically thin semiconductors, can interact with one another with a nonlinearity sufficient to switch an optical probe beam between two discrete logical states. The work, reported in a preprint on May 12 and amplified by ScienceDaily on May 18, 2026, establishes for the first time at room accessible temperatures that the polariton platform supports the nonlinear primitives required for logic operations rather than serving merely as a passive optical interconnect.

The device geometry is built around a high finesse Fabry Perot microcavity into which a monolayer of a transition metal dichalcogenide, in this case tungsten diselenide, is embedded. Strong coupling at room temperature is confirmed by angle resolved photoluminescence showing a Rabi splitting of several tens of millielectronvolts. By tuning the photon to exciton detuning into the lower polariton branch and seeding the cavity with a femtosecond optical pump, the group demonstrates that a second probe beam at a slightly displaced energy can be either transmitted or rejected by the cavity as a function of pump intensity, producing a clean two state nonlinear response.

The energy budget per switching event, extracted from the pump fluence required to produce a measurable polariton density, lands at approximately four femtojoules. That figure is comfortably below the threshold currently considered necessary for photonic primitives to be competitive with state of the art FinFET or gate all around transistors at iso speed, and is two to three orders of magnitude below the energy cost of equivalent operations in current CMOS logic. The team further reports a switching latency in the low picosecond range, consistent with the photonic, rather than excitonic, character of the lower polariton branch in the regime they probe.

Several caveats remain. The current platform requires a precisely fabricated cavity with finesse in the several thousands, a thin film stack whose yield at wafer scale is not yet established, and cryogenic operation for the densest polariton condensates the team explored. Integration with conventional silicon photonics waveguides is unproven, and the demonstrated switch is a single bistable element, not a programmable arithmetic unit. The authors argue, however, that the result provides a quantitatively credible target for downstream engineering and that the same polariton substrate could in principle host analog photonic neural networks, on chip optical reservoir computers, and certain classes of analog quantum simulation. Whether photonic polariton logic ultimately complements silicon transistors at the inference layer of large neural networks, displaces them in narrow accelerator niches, or remains a laboratory curiosity will depend on the next five years of materials, packaging, and systems engineering rather than on the underlying physics, which the Penn paper has now firmly anchored.