Programming the Direction of Heat: A Geometric Pump on a Quantum Processor

A quantum heat pump does not need a conventional piston. In a new experiment reported in an October 6, 2026 preprint, researchers programmed a small loop of control settings into a superconducting quantum processor and detected a heat-transfer component that reverses when the loop is traversed backwards. The working device used one qubit and two auxiliary qubits representing thermal reservoirs.[1][2]

The result is an unusually concrete connection between periodic quantum control and energy management. It is also a lesson in reading quantum-energy headlines carefully: the researchers detected geometric heat pumping, but the processor's noise generated enough additional heating to prevent net cooling.[2] Both facts belong in the headline story.

The experiment separates a programmable heat-pumping signal from unwanted heating. It does not yet turn that signal into a working quantum refrigerator.

For Floquet research, the interest is not a claim of free energy or a broken thermodynamic limit. It is the prospect of using the shape and direction of a repeating control cycle to decide where energy goes—and of testing that prospect on existing quantum hardware.

What was actually demonstrated?

The study, Geometric heat pumping on a quantum processor, implements a driven quantum system on IBM's Heron r2 processor, using the machine named ibm_kingston. A central qubit interacts repeatedly with two helpers, or ancillas. These helpers are prepared in thermal states and stand in for left and right reservoirs at the same temperature.[2]

This is a hardware implementation of a thermal model, not two macroscopic objects whose temperatures are being lowered and raised. The researchers infer the model's exchanged heat from energy measurements of the ancillary qubits immediately before and after their interactions. Those measured changes provide a microscopic calorimetric record: a ledger of energy transferred during the protocol.[2]

The authors report that the measured pumping contribution follows the dependence on cycle period and control-loop shape predicted by exact collision-model simulations, without adjustable parameters. For sufficiently slow driving, it approaches the geometric limit of a continuous open-system description.[1] That is a mechanism-level benchmark, rather than a claim about commercial cooling capacity.

How can heat move between equal-temperature reservoirs?

Equal temperatures remove the usual thermal bias that drives heat from hot to cold. They do not forbid an externally controlled device from moving energy between reservoirs. The important difference is that this experiment supplies a time-dependent control protocol: the system is not simply left alone at equilibrium.

Two properties of the working qubit change during each cycle. One is its energy-level splitting, which determines the energetic cost of an excitation. The other is its orientation relative to the reservoir couplings, which changes how strongly each reservoir can exchange excitations with it.[2]

In the slow-driving picture, the qubit absorbs heat while its splitting decreases and releases heat while the splitting increases. By making those stages preferentially exchange energy with different reservoirs, the protocol produces directional pumping.[2] An everyday analogy is a bucket brigade with timed handoffs: the direction comes from the sequence of operations, not from a permanent slope.

The paper explicitly keeps the energy accounting intact. For a periodic state, the working qubit returns to its initial internal energy after a cycle. At equal reservoir temperatures, the sum of the heat delivered to the reservoirs equals the work supplied by the drive, and that total is non-negative.[2] Moving heat around is not the same as extracting work from one equilibrium temperature.

Why the shape of a loop matters

Imagine drawing a route on a map whose axes are two control settings rather than geographical coordinates. During one cycle, the controls follow a closed curve on that map. In the slow limit, the geometric contribution to transferred heat depends on this curve, not on how quickly particular portions are traversed.[2]

This is the sense in which the pump is geometric. The result is analogous to a Berry-phase effect: completing a closed journey through control space can have a cumulative consequence even though the controls return to their starting values. The authors express the pumped heat through a geometric response and, for equal coupling strengths, through the flux of a curvature over the enclosed region.[2]

Three experimental checks make that abstract description tangible:

• Reverse the journey. Traversing the same loop in the opposite direction reverses the geometric heat contribution.[2]
• Collapse the loop. In the tested family of protocols, eliminating the enclosed area makes the measured orientation-odd heat vanish within error bars.[2]
• Slow the cycle. The measured pumping signal tends toward the predicted adiabatic value for long periods.[2]

These are more informative tests than observing some energy change after applying gates. They connect the measured signal to a specific control mechanism. Importantly, geometric does not automatically mean quantized or topologically protected; this experiment studies a contour-dependent response, not a universal fixed packet of heat for every possible loop.

A tiny reservoir that can also act as a thermometer

Large thermal environments are difficult to represent directly inside a small quantum computer. Collision models take a different approach. A fresh auxiliary system meets the working system briefly, exchanges energy with it, and is then discarded or reset. Repeating these encounters reproduces a controlled form of environmental influence.[2]

In this implementation, an ancilla is reset, prepared with the thermal occupation probabilities, measured, coupled to the working qubit through a partial excitation-exchange operation, and measured again. Resetting it before reuse prevents correlations from being carried into the next encounter. The energy splitting assigned to each ancilla matches the working qubit's instantaneous splitting.[2]

The matching matters because the ideal partial swap conserves the energy of the qubit-ancilla pair during the encounter. An ancilla's energy gain then corresponds to the working qubit's loss. The reservoir helper doubles as a meter, with its before-and-after measurement record providing the exchanged heat.[2]

The driving cycle is divided into 48 cells in the reported implementation. Each cell uses a symmetric left-right-right-left collision ordering to reduce discretization error and bias from letting one reservoir act first. The circuit's initial state is calculated to represent the cycle-periodic regime, avoiding the need to run settling cycles on the processor.[2] These choices show how carefully the abstract thermal machine must be translated into executable gates.

The crucial distinction: subtracting a background is not cooling

The biggest qualification is not hidden in a technical appendix. The authors state that noise-induced heating exceeds the pumped heat in the present implementation, precluding net cooling.[2] A device can contain a real directional pumping mechanism and still warm everything overall if its losses are larger.

To isolate the mechanism, the researchers run the control loop in both directions and take half the difference between the measured heats. The geometric contribution reverses sign, whereas the leading dissipative contribution remains unchanged when the timing profile is reversed. This extracts what the paper calls the orientation-odd component.[2]

The processor's noise background is predominantly orientation-even, so this subtraction removes much of it from the estimator. It does not remove that energy from the physical system. A smaller orientation-odd noise component also remains and attenuates the pumping signal.[2]

A cleaner measurement of a heat pump is not automatically a cleaner-running heat pump. Data analysis can expose the useful mechanism while the hardware still dissipates more energy than it transfers usefully.

This distinction should shape how the result is reported. “Geometric pumping detected on a quantum processor” is supported. “A quantum computer has refrigerated itself with geometric control” is not. Nor does the experiment establish an overall energy advantage after including the control electronics, measurement, reset and cryogenic infrastructure.

Where this fits in Floquet engineering

Floquet research broadly studies quantum systems under repeating drives. This experiment belongs to that periodically driven landscape: the working-qubit Hamiltonian and its effective reservoir contacts follow a recurring schedule. Its central result, however, is a slow-driving geometric response, not the discovery of a new Floquet band structure or a light-induced material phase.[2]

That distinction helps connect different parts of the field without conflating them. Floquet materials research often asks which effective interactions a repeated drive can create. Geometric thermal control asks what a repeated route through parameter space does to energy exchange. Both treat time-dependent control as a design tool, but their observables and practical tests differ.

For quantum-energy applications, the proposed lesson is useful: optimize the full cycle, not simply the strongest pulse or fastest gate. The model separates a contour-dependent pumping contribution from speed-dependent dissipation. Slower operation can make the geometric mechanism clearer, while faster operation can increase heating enough that both reservoirs receive heat.[2] Useful cooling power therefore cannot be inferred from the pumped heat per cycle alone.

A realistic path toward quantum refrigeration

Quantum refrigeration already has a separate experimental benchmark. A peer-reviewed 2025 Nature Physics study, Thermally driven quantum refrigerator autonomously resets a superconducting qubit, demonstrated cooling a transmon below the temperature achievable with any one available bath. That device was powered by a thermal gradient and operated without external feedback.[4]

The October preprint explores a different route: programmable cyclic control with digitally emulated reservoirs. It should not be described as the first quantum refrigerator, or as having already matched the autonomous refrigerator's useful cooling result. Its contribution is the experimentally accessible geometric pumping protocol and its heat-measurement method.[2][4]

What would make the next result stronger?

The authors identify shorter measurement-and-reset waiting times, better gates and longer coherence times as promising ways to reduce physical heating. Their noise simulations highlight errors and relaxation or dephasing, including the working qubit's vulnerability while ancillas are measured and reset.[2] Building on those findings, a practical evaluation should ask:

• Does one reservoir actually lose net heat? Report total heat as well as the background-subtracted pumping component.
• What is the useful cooling rate? Include the full physical duration of preparation, collisions, measurement and reset.
• How much control energy is required? Distinguish the simulated thermal-machine balance from the whole experimental apparatus.
• Does the advantage survive realistic imperfections? Test different circuits, noise conditions and finite temperature differences rather than relying on a single operating point.

These are proposed evaluation criteria, not achievements claimed for the present experiment. They mark the distance between verifying a mechanism and making it useful.

The bottom line

The new preprint offers a concrete answer to a foundational question: can a programmed loop of quantum controls leave a measurable, direction-sensitive signature in heat exchange? On the reported superconducting processor, the answer is yes.[1][2]

The harder engineering question remains open: can the pumping exceed the device's own unwanted heating? This experiment does not yet cross that threshold.[2] Its value is that it makes both the desired process and the obstacle measurable. For a research field connecting Floquet control to practical energy management, that is a more useful milestone than an unsupported promise of beyond-Carnot performance.

Research status: The central result is an experimental report in an arXiv preprint submitted October 6, 2026, not a journal publication verified for this article. Coverage checked October 9, 2026.[1]