Thermodynamic Reservoir Computing

• Thermodynamic Reservoir Computing is an approach that treats reservoir computing as a physical thermodynamic process, integrating energy, entropy, and dissipation constraints.
• It employs conventional input–reservoir–readout architectures in diverse substrates such as quantum many-body systems, frustrated magnets, and ASICs to achieve high-dimensional mapping and spectral filtering.
• This research program quantifies the trade-off between computational performance and energetic cost using metrics like Holevo capacity, irreversible work, and Landauer bounds.

Searching arXiv for papers on thermodynamic reservoir computing and closely related physical/quantum reservoir computing. Thermodynamic Reservoir Computing (TRC) denotes an emerging perspective in which reservoir computing is treated as a physical thermodynamic process with explicit energetic, entropic, and dissipation constraints rather than only as a nonlinear dynamical mapping. In the available literature, this perspective appears in several closely related forms: a non-equilibrium thermodynamic theory of quantum reservoir computing that links predictive performance to free-energy cost, irreversible work, and heat dissipation; a spintronic physical-reservoir framework in which thermally fluctuating frustrated magnets are exploited through frequency-domain filtering; and a theoretical proposal to repurpose voltage-stressed Bitcoin mining ASICs as dissipative physical reservoirs whose timing dynamics may be computationally useful (Ding et al., 2 Jul 2026, Kobayashi et al., 2023, Lafuente et al., 5 Jan 2026). Taken together, these works portray TRC as a research program centered on the proposition that memory, prediction, and readout in reservoir systems are inseparable from the thermodynamics of the underlying substrate.

1. Conceptual scope and relation to reservoir computing

TRC inherits the standard reservoir-computing architecture in which a fixed reservoir transforms inputs into a high-dimensional internal state space and only the readout is trained. The ASIC proposal states this lineage explicitly, identifying high dimensionality, nonlinearity, fading memory or echo-state-like behavior, and the separation property as the relevant desiderata for a physical reservoir (Lafuente et al., 5 Jan 2026). The spintronic work likewise adopts the conventional input–reservoir–linear-readout structure, but situates it in a thermally fluctuating magnetic medium rather than in a simulated recurrent network (Kobayashi et al., 2023).

What distinguishes TRC from a generic physical-reservoir perspective is the insistence that computational utility be analyzed together with energetic cost. The quantum framework makes this most explicit by arguing that quantum reservoir computing is not just a dynamical computation model but a thermodynamic information engine, and by connecting Holevo information capacities, open quantum system dynamics, geometric information measures, quantum dissipation and Landauer limits, and criticality and resonance as the microscopic origin of computational peaks (Ding et al., 2 Jul 2026). In that formulation, the relevant question is not only whether a reservoir can separate temporal inputs, but also what work and dissipation are required to sustain the corresponding state differentiation.

A recurring theme across the literature is that thermodynamic fluctuations are not treated solely as parasitic noise. In frustrated magnets, the useful states are selected from frequency bands that remain robust under thermal agitation (Kobayashi et al., 2023). In the ASIC proposal, voltage stress and thermal fluctuations are hypothesized to enrich the reservoir dynamics near the edge of stability (Lafuente et al., 5 Jan 2026). In the quantum theory, coherence can be either a thermodynamic resource or dissipative overhead depending on whether it contributes to prediction (Ding et al., 2 Jul 2026). This suggests that TRC is best understood not as the elimination of noise, but as the selective exploitation, filtering, or accounting of physical fluctuations in a reservoir substrate.

2. Thermodynamic formalism in quantum reservoir computing

The most developed formalism presently associated with TRC is the non-equilibrium framework for quantum reservoir computing. There the reservoir is modeled as a strongly interacting many-body open quantum system driven by a time-dependent classical input sequence $s_n$, with instantaneous Hamiltonian

$H_{t_n} = H_0 + \lambda s_n H_1,$

where $H_0$ is the intrinsic many-body Hamiltonian, $H_1$ is the input-coupling operator, $\lambda \ll 1$ is the weak-driving parameter, and $s_n$ is a zero-mean stochastic input signal (Ding et al., 2 Jul 2026). The temporal protocol is sequential: a signal is injected, the reservoir evolves coherently under $H_{t_n}$, it relaxes thermally toward the Gibbs state, and the resulting conditional state encodes task-relevant information.

Within this framework, the paper distinguishes memory states $\rho^{\rm m}_{s_n}$, conditioned on the present or past input, from predictive states $\rho^{\rm p}_{s_{n+1}}$, conditioned on the future target. These ensembles share the same unconditional average state $\bar{\rho}_{t_{n+1}}$, which permits a controlled comparison of memory and prediction on equal footing. Their information-processing capability is quantified by Holevo capacities,

$H_{t_n} = H_0 + \lambda s_n H_1,$0

applied separately to memory capacity $H_{t_n} = H_0 + \lambda s_n H_1,$1 and predictive capacity $H_{t_n} = H_0 + \lambda s_n H_1,$2 (Ding et al., 2 Jul 2026).

A central analytical step is the weak-driving expansion around the baseline state,

$H_{t_n} = H_0 + \lambda s_n H_1,$3

under which the relative entropy becomes, to second order,

$H_{t_n} = H_0 + \lambda s_n H_1,$4

with the Bogoliubov–Kubo–Mori metric

$H_{t_n} = H_0 + \lambda s_n H_1,$5

This yields a geometric interpretation in which capacity is a local curvature of state space fixed by the thermal equilibrium structure (Ding et al., 2 Jul 2026).

The thermodynamic cost side is formulated through quantum relative entropy, non-equilibrium free energy, irreversible work, heat dissipation, and Landauer-like bounds. The non-equilibrium free energy is

$H_{t_n} = H_0 + \lambda s_n H_1,$6

and the irreversible work is

$H_{t_n} = H_0 + \lambda s_n H_1,$7

For the full temporal process, the theory derives

$H_{t_n} = H_0 + \lambda s_n H_1,$8

together with the generalized Landauer bound

$H_{t_n} = H_0 + \lambda s_n H_1,$9

Here $H_0$0 and $H_0$1 denote classical and quantum informational dissipation, while $H_0$2 and $H_0$3 are the corresponding classical and quantum components of the Landauer cost for continuous temporal processing (Ding et al., 2 Jul 2026).

The same work uses the coherence decomposition

$H_0$4

to separate classical population information from the coherence contribution. This decomposition is fundamental for TRC because it identifies which part of the reservoir’s internal differentiation is attributable to populations and which to quantum superposition. It also underpins the definition of quantum informational dissipation as the coherence retained for memory but not useful for prediction (Ding et al., 2 Jul 2026).

3. Criticality, resonance, and frequency-domain computation

In the quantum formulation, computational peaks occur near quantum critical regions. The microscopic explanation combines energy-gap closing near a continuous phase transition, a dense low-energy spectrum, and resonance between reservoir transition frequencies and the input frequency. The transition frequencies are

$H_0$5

and the relevant dynamic factor is a filtered temporal accumulation term $H_0$6 that behaves like a resonance response to the input’s characteristic frequency $H_0$7. When

$H_0$8

constructive interference is obtained and capacity is maximized; far from resonance, oscillatory terms destructively interfere and suppress capacity. As the system approaches criticality, the gap closes, $H_0$9, and the dense low-energy spectrum allows transition frequencies to sweep through the low-frequency content of the input (Ding et al., 2 Jul 2026). In this account, criticality supplies the transition structure, resonance matches reservoir frequencies to temporal input frequencies, and capacity peaks where those align.

A closely related spectral logic appears in the spintronic reservoir based on frustrated magnets. There, information is encoded into AC magnetic fields and propagated through a classical antiferromagnetic Heisenberg model on a triangular lattice,

$H_1$0

with dynamics governed by the stochastic Landau–Lifshitz–Gilbert equation including thermal noise (Kobayashi et al., 2023). The key claim is that short-term memory is stored in the spin dynamics at frequencies around that of the AC field and can be preserved against thermal noise by filtering out irrelevant signals at other frequencies. The reservoir state is constructed from time samples of $H_1$1, and the output is a linear readout,

$H_1$2

with performance evaluated by the determination coefficient $H_1$3 (Kobayashi et al., 2023).

The spintronic paper extends this spectral viewpoint to frequency-division multiplexing. Distinct input streams are assigned different frequencies; the spectra then exhibit distinct peaks at those frequencies, and each stream can be recovered with the appropriate filter. Exchange interactions further create second harmonics, sum frequencies, and difference frequencies, thereby enabling information processing among different frequency threads (Kobayashi et al., 2023). A plausible implication is that TRC often depends less on using the entire raw state space than on isolating the thermodynamically stable, resonant, or task-relevant dynamical modes of that state space.

4. Representative substrates and architectures

The current literature presents TRC through markedly different physical substrates, ranging from quantum many-body systems to classical spin systems and voltage-stressed digital ASICs.

Work	Physical substrate	TRC emphasis
(Ding et al., 2 Jul 2026)	Strongly interacting many-body open quantum system	Holevo capacity, BKM geometry, dissipation, Landauer bound, critical resonance
(Kobayashi et al., 2023)	Frustrated magnet on a triangular lattice	Frequency filtering, thermally robust STM, frequency-division multiplexing, spatial parallelization
(Lafuente et al., 5 Jan 2026)	BM1366 Bitcoin mining ASIC	Timing dynamics, thermodynamic noise, edge-of-stability operation, CHIMERA instrumentation

The quantum substrate is explicitly a driven open quantum system with conditional memory and predictive ensembles, and its principal contribution to TRC is analytical: it establishes a thermodynamic theory linking predictive performance to microscopic energetic cost (Ding et al., 2 Jul 2026).

The frustrated-magnet substrate is a spintronic physical reservoir in which the computational variables are the time-dependent spin dynamics under stochastic Landau–Lifshitz–Gilbert evolution. It is presented as suitable for reservoir computing because it offers nonlinearity from spin interactions, high-dimensional dynamics from many interacting spins, short-term memory, physical tunability via magnetic fields and read-out selection, and potential integration even down to the level of a single spin. The paper also identifies experimentally plausible triangular-lattice materials including NaCrO$H_1$4, LiCrO$H_1$5, and Ba$H_1$6CoSb$H_1$7O$H_1$8 (Kobayashi et al., 2023).

The ASIC proposal introduces Holographic Reservoir Computing (HRC) and the CHIMERA system architecture. In this framework, the SHA-256 pipeline is not treated as the reservoir’s computation in the usual algorithmic sense but as a deterministic nonlinear diffusion operator embedded in a noisy physical substrate, with observable computation arising from timing dynamics rather than hash outputs themselves. CHIMERA is organized into three layers: The Ghost, which polls telemetry, sends voltage and frequency commands via the AxeOS firmware API, timestamps share events, and collects voltage, temperature, power, and hashrate; The Muse, which extracts inter-arrival time statistics, coefficient of variation, histogram entropy, Hamming distance between consecutive hashes, and phase-amplitude encoding; and The Sentinel, which provides PID thermal control, voltage limits, anomaly detection, and logging (Lafuente et al., 5 Jan 2026).

The BM1366 hardware is described as having 138 hash cores, a 300–500 MHz operating range, an 850–990 mV core voltage range, and nominal power roughly 5–15 W. The experimental program uses voltage stress and frequency sweeps, with telemetry polling every 3 s and dwell times of 60 s for voltage and 30–45 s for frequency (Lafuente et al., 5 Jan 2026). In TRC terms, this is not yet a validated reservoir but a substrate proposal equipped with a concrete control and measurement stack.

5. Energetic trade-offs, coherence, and thermal robustness

A defining claim of TRC is that better temporal processing is not thermodynamically free. In the quantum theory, predictive and memory capacities are geometric measures of distinguishability; distinguishability requires work and dissipation; and irreversible work bounds the heat dissipation. The resulting trade-off is explicit: the critical resonance that unlocks optimal predictive capacity inherently maximizes informational dissipation and the irreversible work required for environmental erasure (Ding et al., 2 Jul 2026). This statement places computational optimality and thermodynamic cost in direct tension rather than treating energy efficiency as an afterthought.

Quantum coherence occupies a particularly nuanced role. Through the decomposition $H_1$9 and $\lambda \ll 1$0, coherence can be identified as a distinct contribution to predictive capacity. If coherences are created but do not help future prediction, they contribute to extra dissipation. If they encode temporally useful structure in a non-Markovian drive, they can lower the effective Landauer bound. For Markovian driving, coherent dissipation is non-negative; for non-Markovian driving, it need not be, so coherent dissipation can become negative and act as a thermodynamic resource (Ding et al., 2 Jul 2026). This directly counters the common simplification that coherence is always either beneficial or wasteful.

The spintronic work presents an analogous trade-off in classical thermodynamic language. Thermal noise ordinarily destroys the short-term memory on which reservoir computing relies, but memory retained around the input frequency is resistant to thermal fluctuations, whereas other frequencies act as noise. Without frequency filtering, memory older than $\lambda \ll 1$1 quickly disappears even at extremely low $\lambda \ll 1$2; with filtering, short-term memory is preserved much better and remains accurate up to larger delays. Increasing the input amplitude $\lambda \ll 1$3 broadens the usable temperature range by making the signal peaks stand out above thermal noise (Kobayashi et al., 2023). In this setting, efficiency is achieved not by removing thermal fluctuations globally, but by preserving the spectral components that remain informative under those fluctuations.

The ASIC proposal advances a different kind of efficiency claim. It argues, as a theoretical projection based on a Hierarchical Number System representation, that conventional von Neumann systems scale as $\lambda \ll 1$4 and more concretely $\lambda \ll 1$5, whereas the HNS-based CHIMERA representation scales as $\lambda \ll 1$6 and $\lambda \ll 1$7. The ratio is written as approximately

$\lambda \ll 1$8

and for typical reservoir dimensions such as $\lambda \ll 1$9, the paper projects an improvement on the order of 10,000×. However, the same paper repeatedly states that this is a theoretical upper bound, not a measured result, and does not provide a proof that the ASIC reservoir realizes this scaling in practice (Lafuente et al., 5 Jan 2026). That caveat is central to any neutral appraisal of TRC efficiency claims.

6. Validation status, misconceptions, and unresolved problems

The three strands of TRC literature are not at the same stage of validation. The quantum paper provides a theoretical framework and analytical results, including the generalized Landauer statement for continuous temporal processing and the identification of critical resonance as the origin of computational peaks, but it is primarily a theory of energetic limits and design principles for quantum learning devices and energy-efficient quantum neuromorphic hardware (Ding et al., 2 Jul 2026). It should not be conflated with an empirical benchmark study of deployed hardware.

The spintronic work is more concrete as a physical-reservoir demonstration, but it also identifies clear limitations. The reservoir is a simple triangular-lattice Heisenberg model with relatively low-dimensional internal structure. Tasks such as XOR and XNOR are hard when using only linear-frequency channels and become effective only when relying on nonlinear difference- and sum-frequency components. Those nonlinear higher-harmonic channels are vulnerable to thermal noise, making linearly inseparable gates nearly infeasible at finite temperature in this model; stronger or different interactions are suggested as a route to improvement (Kobayashi et al., 2023). Thus, even when TRC behavior is physically instantiated, thermal robustness is channel-dependent and task-dependent.

The ASIC proposal is the most speculative. It is explicit that the work is a theoretical proposal with preliminary instrumentation and exploratory observation, not a validated demonstration of Thermodynamic Reservoir Computing. Reported observations include non-Poissonian variability, coefficient-of-variation excursions above and below 1, and systematic changes as voltage drops toward the stability boundary, but sample sizes are small, confounds are not controlled, multi-chip reproducibility is missing, statistical significance analysis is pending, and no FFT/PSD analysis, invalid-share counting, or direct hardware probing has yet been completed (Lafuente et al., 5 Jan 2026). The proposed “Silicon Heartbeat,” sometimes described informally as around 2.4 Hz, is explicitly not yet validated. Likewise, necessary RC properties such as echo-state-like fading memory and separation are identified as required but not shown, and no benchmark tasks such as NARMA-10 or Mackey–Glass are reported (Lafuente et al., 5 Jan 2026).

Several misconceptions can therefore be excluded. TRC is not equivalent to any use of a physical substrate for computation; the literature emphasizes readout structure, temporal state evolution, and substrate-specific memory mechanisms. TRC does not imply that noise is always useful; the spintronic work distinguishes thermally resilient frequency channels from noise-dominated ones, and the quantum work distinguishes useful coherence from coherence that counts as dissipation (Kobayashi et al., 2023, Ding et al., 2 Jul 2026). Nor does TRC presently denote a single standardized platform. The term spans a rigorous quantum thermodynamic theory, a concrete spintronic design principle, and a pre-validation ASIC substrate proposal (Ding et al., 2 Jul 2026, Kobayashi et al., 2023, Lafuente et al., 5 Jan 2026).

From these works, a coherent research agenda nonetheless emerges. It includes establishing fading memory and separability in proposed substrates, determining whether observed timing structure originates in silicon physics or in software and network artifacts, validating spectral signatures such as the putative Silicon Heartbeat, measuring actual error rates rather than diffusion proxies in ASIC-based systems, and designing reservoirs that preserve only predictively useful coherences or thermally stable frequency channels (Lafuente et al., 5 Jan 2026, Ding et al., 2 Jul 2026, Kobayashi et al., 2023). A plausible implication is that mature TRC will require joint optimization over substrate dynamics, spectral selectivity, and thermodynamic cost, rather than optimization over prediction accuracy alone.