Probabilistic computing utilizing HfO2-based stochastic ferroelectric tunnel junctions Authors: Zeyu Guan, Hansheng Zhu, Yaoxin Li, Yuanzhenzi Lu, Haifeng Bu, Bo Xu, Zhengxu Zhu, Shengchun Shen, Yuewei Yin, Xiaoguang Li Published: 2026 ABSTRACT Probabilistic neural networks are good at solving complex optimization tasks, but require stochastic, energy-efficient probabilistic bit (p-bit) neurons and reliable artificial synapses. Here we show stochastic ferroelectric tunnel junctions (s-FTJs) and reliable-FTJs (r-FTJs) by tuning the oxygen vacancy concentration in Hf0.5Zr0.5O2 ferroelectric film, which are utilized to set up p-bit neurons and synapses, respectively. The s-FTJ-based p-bit outputs 0 or 1 with a tunable probability, and it can operate as a true random number generator at a probability of 0.5. The write power per p-bit is ~76 nW, significantly lower than other reported p-bit implementations. A hardware prototype of a four-neuron Boltzmann machine is experimentally constructed for probabilistic computing, which successfully solves the maximum independent set problem. Simulations show that a 655-neuron Boltzmann machine can accurately predict the secondary structure of a 64-nucleotide RNA. This work provides a high-performance probabilistic computing solution with low energy consumption and excellent process compatibility. Introduction Probabilistic computing, which utilizes models like Boltzmann machines, Bayesian networks, and Ising models, provides significant advantages in tasks involving uncertainty quantification, generative modelling, and combinatorial optimization1. Representative issues, such as integer factorization, Max-Cut, the travelling salesman problem (TSP), and the maximum independent set (MIS) problem, can be mapped onto the search for low-energy states within these models. Probabilistic computing is intrinsically suited to solve such problems for which classical algorithms (e.g., dynamic programming and depth-first search) are not as effective or naturally adapted2. As illustrated in Fig. 1a, probabilistic neural networks (using a Boltzmann machine as an example) comprise probabilistic neurons (i.e., p-bits) whose binary states (outputs) are denoted as xi (= 0 or 1), with synapses (weights \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${w}_{{ij}}$$\end{document}wij) connecting neurons i and j. Near-optimal solutions (i.e., the states of all neurons) are obtained by iteratively updating neuron states to minimize the network energy3, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E=\mathop{\sum }_{{ij}}{w}_{{ij}}{x}_{i}{x}_{j}-\mathop{\sum }_{i}{\theta }_{i}{x}_{i}$$\end{document}E=∑ijwijxixj−∑iθixi, where constant term \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\theta }_{i}$$\end{document}θi is the bias of neuron xi. The update of neuron states is guided by the local energy gradient, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${z}_{i}=-\partial E/\partial {x}_{i}=-\mathop{\sum }_{{ij}}{w}_{{ij}}{x}_{j}{+\theta }_{i}$$\end{document}zi=−∂E/∂xi=−∑ijwijxj+θi. Computation in a Boltzmann machine thus involves two principal steps: (1) a synapse operation: calculating \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${z}_{i}$$\end{document}zi via the synapse weight \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${w}_{{ij}}$$\end{document}wij and neuron state xi; and (2) a p-bit operation: stochastically setting xi = 0 or 1 according to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${z}_{i}$$\end{document}zi fed back from other neurons2. The output probability of xi = 1 is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\sigma} ({z}_{i})$$\end{document}σ(zi), where σ is the sigmoid function. In other words, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${x}_{i}=u({\sigma} ({z}_{i})-r)$$\end{document}xi=u(σ(zi)−r), where u is the unit step function and r (0 < r < 1) is a uniformly distributed stochastic number.Fig. 1Schematic diagram of a Boltzmann machine based on FTJs.a Network topology of the Boltzmann machine and a schematic of solving combinatorial optimization problems. b Hardware implementation of the Boltzmann machine using s-FTJs and r-FTJs. Schematic diagram of a Boltzmann machine based on FTJs. a Network topology of the Boltzmann machine and a schematic of solving combinatorial optimization problems. b Hardware implementation of the Boltzmann machine using s-FTJs and r-FTJs. Several approaches have been explored to implement stochastic p-bits at the hardware level. Among them, CMOS-based solutions typically rely on complex random number generation circuits, requiring thousands of transistors per p-bit and leading to large chip area and high power consumption4–7. Recently, several emerging nanodevices such as magnetic tunnel junctions (MTJs)8–10, CuTe-based diffusion memristors3, NbOx oscillating memristors11, and VO2 Mott resistive switching devices12, have provided alternative physical realization of p-bits by employing thermal fluctuations or defect-induced stochasticity. However, these devices suffer from relatively high power consumption, and most prior studies focus solely on the construction of p-bit unit, while the synapse operation in probabilistic neural networks still relies on conventional digital processors, limiting the energy efficiency at the system-level. Compared to the aforementioned nanodevices, ferroelectric devices such as ferroelectric tunnel junctions (FTJs) and ferroelectric field-effect transistors (FeFETs) exhibit significant advantages in terms of energy efficiency13–18. Recently, the design of a p-bit based on a FeFET has been theoretically proposed by exploiting dipole fluctuations induced by thermal noise, but still awaits experimental verification19. As is known, oxygen vacancies in ferroelectric oxide materials strongly affect the polarization dynamics and resistive switching characteristics of devices20–26. For instance, Hernandez-Martin et al. theoretically and experimentally demonstrated that oxygen vacancies in Ag/BaTiO3/La0.7Sr0.3MnO3 FTJs are coupled with ferroelectric domains, enabling controllable sign reversal of electroresistance within a single FTJ23. In particular, higher oxygen‑vacancy concentrations can enhance the stochasticity of resistance switching27. Inspired by the aforementioned results, we propose a strategy to achieve dual-mode functionality in Pt/Hf0.5Zr0.5O2 (HZO)/TiO2/TiN FTJs by adjusting the oxygen vacancy concentration within the HZO ferroelectric layer. Specifically, r-FTJs with reliable resistive switching are constructed for deterministic synapse operations, while s-FTJs with enhanced stochasticity serve as p-bit generators, as illustrated in Fig. 1b. To verify the system performance, we experimentally built a four-neuron Boltzmann machine constructed from four s-FTJs and a 4 × 4 crossbar-structured r-FTJ array, which successfully solved the MIS problem with a low write power of ~76 nW per p-bit. Furthermore, simulations expanded the system to a Boltzmann machine with 655 neurons, accurately predicting the secondary structure of an RNA composed of 64 nucleotides. Compared to reported p-bit implementations based on emerging devices, our approach exhibits significant advantages in terms of energy efficiency and fabrication simplicity. Results and discussion Structural and electrical characterizations To realize the dual-mode functionality required for the Boltzmann machine, we adjusted the purge duration of the Hf/Zr precursors during atomic layer deposition (ALD), thereby manipulating the oxygen vacancy concentration in the HZO layer. Both s-FTJs (20 s purge) and r-FTJs (40 s purge) have the same Pt/HZO/TiO2/TiN stacking structure (Supplementary Fig. 1). Cross-sectional high-resolution transmission electron microscopy (HRTEM) images (Fig. 2a, b) reveal clear interfaces and high crystallinity for both devices. The results of HRTEM and quantitative X-ray energy dispersive spectroscopy (EDS, Supplementary Fig. 2) confirm HZO and TiO2 thicknesses of ~3.5 nm and ~1.5 nm, respectively, in both types of FTJs. The TiO2 interlayer imposes in-plane tensile stress on HZO to promote tetragonal-to-orthorhombic (t-to-o) phase transition during annealing28–30, and simultaneously acts as an oxygen reservoir to mitigate oxygen vacancy formation in HZO31. Furthermore, it isolates the HZO film from the TiN bottom electrode, thus suppressing interfacial dead layer formation32.Fig. 2Structural characterization of Pt/HZO/TiO2/TiN.a, b Cross-sectional HRTEM images. Scale bar: 10 nm. c o/t-phase proportion. d Grain size statistics obtained from SEM. e, f X-ray photoelectron spectroscopy spectra of Hf 4f and multi-peak Gaussian fitting results. Structural characterization of Pt/HZO/TiO2/TiN. a, b Cross-sectional HRTEM images. Scale bar: 10 nm. c o/t-phase proportion. d Grain size statistics obtained from SEM. e, f X-ray photoelectron spectroscopy spectra of Hf 4f and multi-peak Gaussian fitting results. The crystalline phases observed in the magnified image of the HZO layers in Fig. 2a, b are identified as the o-phase and t-phase, and the o-phase is widely recognized as the ferroelectric phase in HZO films33. Phase proportion analyses of 10 HRTEM images per sample (Supplementary Fig. 3) indicate that both FTJs are dominated by the ferroelectric o-phase, with the s-FTJ exhibiting a slightly lower o-phase proportion (62%, Fig. 2c). Grazing-incidence X-ray diffraction (GIXRD) multi-peak Gaussian fitting results (Supplementary Fig. 4) corroborate these results. The higher t-phase proportion in the s-FTJ should be related to its smaller grains favoring t-phase stabilization33,34, as quantified in Fig. 2d and Supplementary Note 5. Compared to the s-FTJ (average grain size (avg.): ~26.1 nm, standard deviation (s.d.): ~10.4 nm, s.d./avg.: ~40%), the r-FTJ exhibits a larger average grain size (avg.: ~32.6 nm, s.d.: ~11.5 nm) and better grain uniformity (as indicated by a smaller s.d./avg. of ~35%), observed both in plan-view HZO/TiO2/TiN stacks (Fig. 2d and Supplementary Fig. 5) and cross-sectional profiles (Supplementary Fig. 6). Furthermore, a shorter purge during ALD growth leaves more carbon residues in the s-FTJ, as depicted by XPS atom-content data (Supplementary Table 1)35. Hf 4f spectra (Fig. 2e, f) display Hf(4-δ)+ components demonstrating the existence of oxygen vacancies, and multi-peak Gaussian fitting results show nearly twice the oxygen vacancy concentration in the s-FTJs compared with the r-FTJs, consistent with the higher carbon incorporation36. Multi-peak Gaussian fitting results of O 1s spectra (Supplementary Fig. 7) quantifying non-lattice oxygen content also support this conclusion37,38. Figure 3a, b present ten successive resistance versus pulsed voltage (Vp, pulse width = 1 µs) hysteresis loops, measured at a read voltage (Vread) of 0.5 V, for the s‑FTJ and r‑FTJ, respectively. The statistical distributions of switching voltages extracted from these loops are plotted in the lower panels. For both FTJs, the switching voltages exhibit certain distributions, related to the stochastic nature of polarization reversal when an electric field comparable to the switching voltage is applied to the device39,40. First-principles calculations show that HZO ferroelectric domains may switch through several paths with energy barriers as low as a few tens of meV41,42, making them perturbable by thermal fluctuations with the aid of electric field43. In comparison with the r-FTJ, the s-FTJ displays higher average switching voltages and larger standard deviations, which is consistent with the piezoresponse force microscopy (PFM) results in Supplementary Fig. 8. The differences in switching voltage fluctuations between the two FTJs can be attributed to the ferroelectric domain switching dynamics sensitive to oxygen vacancies, because a higher concentration of oxygen vacancies acts as pinning centers, making the switching process highly stochastic and dependent on the local defect landscape44,45. Both types of FTJs exhibit a clear asymmetric switching. The set transition is relatively sharp with lower set voltage in a narrow voltage window, whereas the reset-side switching is gradual and more stochastic with a relatively broader reset voltage range, particularly in the s‑FTJ. This behavior can be attributed to the asymmetric electrodes, the TiO2 depletion/interfacial pinning effect, and a nonuniform distribution of oxygen vacancies (Supplementary Note 8).Fig. 3Resistive switching characteristics of both types of FTJs.Ten consecutive resistance-voltage (R-Vp) hysteresis loops with a pulse width of 1 µs for a the s-FTJ and b the r-FTJ. c Resistance switching between HRS and sub-HRS under repeated cycling. The rightmost subplots display the corresponding statistical distribution of sub-HRS plotted for both types of FTJs. The avg. and s.d. are annotated for the switching voltage distributions in (a, b) and the sub-HRS resistance distributions in (c). Resistive switching characteristics of both types of FTJs. Ten consecutive resistance-voltage (R-Vp) hysteresis loops with a pulse width of 1 µs for a the s-FTJ and b the r-FTJ. c Resistance switching between HRS and sub-HRS under repeated cycling. The rightmost subplots display the corresponding statistical distribution of sub-HRS plotted for both types of FTJs. The avg. and s.d. are annotated for the switching voltage distributions in (a, b) and the sub-HRS resistance distributions in (c). Highly related to the switching voltage fluctuation, the intermediate resistance states in the R-Vp loops exhibit pronounced cycle-to-cycle instability, while the high (HRS) and low resistance states (LRS) show much better repeatability, as illustrated in Fig. 3a, b. To further demonstrate the stochasticity of intermediate resistance state, Fig. 3c shows the periodic switching between reliable HRS (after applying a large enough positive voltage pulse to fully switch ferroelectric polarization pointing toward the bottom TiN electrode, defined as the reset voltage) and stochastic sub-HRS (after applying a negative programming pulse close to the switching voltage, defined as the partial-set voltage). Owing to their different switching voltages, the reset/partial-set voltages were +2.2 V/−1.15 V for s-FTJs and +1.85 V/−1 V for r-FTJs, respectively, to realize HRS and sub-HRS with similar amplitudes. To quantify the resistance fluctuation, we define the resistance fluctuation ratio (RFR) as the relative standard deviation of the measured resistance. It is observed that the RFRs are ~18% and ~3.7% for s-FTJ and r-FTJ, respectively. The higher RFR in s-FTJ should be related to its higher oxygen vacancy concentration, because when an electric field is applied, oxygen vacancy migration will couple with ferroelectric domain switching and enhance the stochasticity of resistance switchings22–24. After applying a large positive reset pulse voltage, the ferroelectric polarization switches fully and deterministically, and meanwhile, positively charged oxygen vacancies tend to accumulate near the TiO2/TiN interface, where their migration is strongly suppressed due to the high activation energy barriers in TiN46–48. These result in relatively reliable resistance with small fluctuations at HRS. However, the sub-HRS was obtained after applying a negative partial-set pulse voltage. The polarization switches partially and stochastically because the partial-set voltage is close to the switching voltage40, and oxygen vacancies driven by this pulse are injected into the HZO layer, where lower migration barriers give them greater spatial distribution freedom49,50. These lead to the enhanced resistance fluctuations at sub-HRS, especially for s-FTJs with more oxygen vacancies (Fig. 3c). The endurance cycling was evaluated for both FTJs, between HRS and sub-HRS, as well as between HRS and LRS (Supplementary Fig. 9). In addition, the endurance between sub-HRS and LRS was also characterized for the s‑FTJ (Supplementary Fig. 10). On the one hand, under identical voltage conditions (+2.2 V/−1.15 V at 1 µs), the r-FTJ not only survives more cycles but also maintains more stable resistance switching with a larger switching ratio than the s-FTJ. On the other hand, for both FTJs, cycling between HRS and sub-HRS endures more cycles than that between HRS and LRS or between sub-HRS and LRS. This motivates the use of the HRS‑to‑sub‑HRS operating mode for the Boltzmann machine, as it improves device lifetime and simultaneously enables the reproducible stochastic behavior. Notably, MΩ-level high resistances (HRS and sub-HRS) are dominated by ferroelectric tunneling transport, suggesting that oxygen vacancies primarily serve to modulate the ferroelectric behavior rather than forming conductive filaments. Single probabilistic bit implementation To implement p-bits with the s-FTJs, we designed a simple circuit shown in Fig. 4a, consisting of an s-FTJ, a 5 MΩ standard resistor, and a comparator. The s-FTJ is driven by the drive voltage Vd illustrated in Supplementary Fig. 11, which includes reset (+2.4 V, pulse duration 1 μs), partial-set (−1.2 V, pulse duration 1 μs), and read (+1 V, pulse duration 540 μs) stages. Here, the read pulse of 540 μs is used, which is mainly limited by the delay of the external readout circuitry (Supplementary Fig. 12). A pair of reset and partial-set pulses stochastically modulates the s-FTJ resistance around 5 MΩ at sub-HRS. Consequently, during the read stage the node voltage Vn,i between the s-FTJ and the 5 MΩ resistor follows:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${V}_{{{\rm{n}}},i}={V}_{{{\rm{d}}},{{\rm{read}}}}\times \frac{{R}_{{{\rm{resistor}}}}}{{R}_{{{\rm{s}}}-{{\rm{FTJ}}},i}+{R}_{{{\rm{resistor}}}}}$$\end{document}Vn,i=Vd,read×RresistorRs−FTJ,i+Rresistorwhere \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${V}_{{{\rm{d}}},{{\rm{read}}}}$$\end{document}Vd,read, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{{{\rm{s}}}-{{\rm{FTJ}}},i}$$\end{document}Rs−FTJ,i, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{{{\rm{resistor}}}}$$\end{document}Rresistor are Vd at read-stage, resistances of the s-FTJ and resistor, respectively, and i denotes the p-bit index. A fluctuated \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{{{\rm{s}}}-{{\rm{FTJ}}},i}$$\end{document}Rs−FTJ,i from 4 to 6 MΩ (Fig. 3c) leads to a Vn,i fluctuation of ~100 mV, which is sufficient to construct the p-bit circuit8.Fig. 4Design and experimental results of a single p-bit.a Electrical schematic of a p-bit circuit. b Time traces of Vo,i for three representative Vref,i values. c Probability of Vo,i = 2 V versus Vref,i. Error bars indicate the standard deviation. Solid line is the fitting result by Eq. (2). d Power consumption for generating a single p-bit by different p-bit circuit implementations. Design and experimental results of a single p-bit. a Electrical schematic of a p-bit circuit. b Time traces of Vo,i for three representative Vref,i values. c Probability of Vo,i = 2 V versus Vref,i. Error bars indicate the standard deviation. Solid line is the fitting result by Eq. (2). d Power consumption for generating a single p-bit by different p-bit circuit implementations. This fluctuated Vn,i is fed to the negative input of the comparator, where it is compared in real-time with the reference voltage (Vref,i, ranging from 0.2 V to 0.8 V) at the positive input. When Vref,i > Vn,i, the comparator outputs Vo,i = 2 V, and when Vref,i < Vn,i, it outputs Vo,i = 0 V. Figure 4b shows the time traces of measured Vo,i for Vref,i = 0.48, 0.51, and 0.53 V, demonstrating the tunable probability (Pi) for outputting Vo,i = 2 V. Each experiment at a given Vref,i consists of 500 sampling cycles, with each cycle lasting 600 μs (i.e., with Vd completing a full cycle from reset, partial-set to read stages). To quantify the relationship between output probability and Vref,i, we selected 10 different Vref,i values and performed three independent tests for each. The statistical results are shown in Fig. 4c, where each data point represents the probability of outputting 2 V. The small standard deviation reflects that the stochastic behavior of s-FTJs is highly repeatable. The data align well with the sigmoid function51 in Eq. (2), with fitted parameters ai = 38 and bi = 0.5, thus enabling p-bit construction2. According to Eq. (2), increasing the reference voltage Vref,i from 0.4 to 0.6 V gradually raises the probability of Vo,i = 2 V from ~0 to ~1.2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${P}_{i}=\frac{1}{1+{e}^{-{a}_{i}\cdot \left({V}_{{{\rm{ref}}},i}-{b}_{i}\right)}}$$\end{document}Pi=11+e−ai⋅Vref,i−bi The s-FTJ-based p-bit can also work as a true random number generator (TRNG)40, when the probability of outputting 0 V or 2 V equals 0.552. To verify the feasibility of TRNG operation, a 10,000-bit sequence was generated from a single p-bit operating at the probability of ~0.5 by setting Vref,i to ~0.51 V. The sequence passed the National Institute of Standards and Technology (NIST) SP800-22 Statistical Test Suite53 and exhibited excellent stochastic characteristics (Supplementary Table 2). The high randomness quality stems from the intrinsic stochasticity of ferroelectric domain switching coupled with oxygen vacancy migration. The s-FTJ-based p-bit can be operated with a lower power consumption. As shown in Fig. 4d and Supplementary Note 12, the s-FTJ-based p-bit achieves a low write power of ~76 nW, significantly outperforming other p-bit implementations based on CMOS5, MTJ8,10, and RRAM3,11. Here, the power of comparator is excluded for all implementations to focus on the core stochastic switching element, considering that the on-chip comparator consumes as little as ~nW54. The substantially smaller power of FTJ primarily arises from its low current density. This breakthrough provides a technological pathway for constructing highly energy-efficient probabilistic computing systems. Boltzmann machine with probabilistic bits and synapse array To demonstrate the practical feasibility of an FTJ-based Boltzmann machine, we constructed a hardware system consisting of four physical s-FTJ p-bit devices and a 4 × 4 r-FTJ array. We validated the system by solving an MIS problem55, a classical combinatorial optimization problem56 that identifies the largest subset of mutually non-adjacent nodes in an undirected graph. For a four-node graph (Supplementary Fig. 14a), the implementation uses four s-FTJs to act as four stochastic neurons (i.e., p-bits), along with a 4 × 4 r-FTJ array that implements the synaptic couplings (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${w}_{{ij}}$$\end{document}wij) between every two of the neurons (xi and xj). The system adopts a modular architecture with two core functional units (Fig. 5a and light-blue box in Supplementary Fig. 14b): (1) a stochastic neuron module, implemented by a p-bit circuit board with four s-FTJs, and (2) a synaptic coupling module based on an FTJ array. And the system is controlled by a field-programmable gate array (FPGA, red box in Supplementary Fig. 14b), which maps the synaptic readout current (IBL,i) to the reference voltage (Vref,i). The multi-channel digital-to-analog and analog-to-digital converter (DAC/ADC) board (blue box in Supplementary Fig. 14b) is utilized to generate pulse voltages and sense current throughout the process. In the stochastic neuron module, each p-bit, driven by a specific periodic voltage sequence (Supplementary Fig. 11), outputs either 0 V or 2 V whose probability is controlled by the reference voltage. Figure 5b plots the output probability of each p-bit as a function of its reference voltage, exhibiting good device-to-device uniformity.Fig. 5Experimental demonstration of four-p-bit Boltzmann machine based on FTJs.a Photograph of the custom printed circuit board containing four p-bit circuits and a 4 × 4 r-FTJ array. b Probability-reference voltage characteristic of four p-bits used in the hardware-implemented Boltzmann machine. c Retention characteristic of the 16 devices in the 4 × 4 r-FTJ array. d Time traces of Vo,i (digitized to 0 or 2 V) for each p-bit from 1000 captures collected over 600 ms. e Probability distribution of all Boltzmann network states. The red bars denote the correct solutions. Error bars represent the standard deviation from three independent experiments. Experimental demonstration of four-p-bit Boltzmann machine based on FTJs. a Photograph of the custom printed circuit board containing four p-bit circuits and a 4 × 4 r-FTJ array. b Probability-reference voltage characteristic of four p-bits used in the hardware-implemented Boltzmann machine. c Retention characteristic of the 16 devices in the 4 × 4 r-FTJ array. d Time traces of Vo,i (digitized to 0 or 2 V) for each p-bit from 1000 captures collected over 600 ms. e Probability distribution of all Boltzmann network states. The red bars denote the correct solutions. Error bars represent the standard deviation from three independent experiments. Simultaneously, the synaptic coupling module receives these stochastic signals and feeds them into the experimental 4 × 4 r-FTJ array to perform synapse operation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathop{\sum }_{{ij}}{w}_{{ij}}{x}_{j}$$\end{document}∑ijwijxj. Each r-FTJ in the array is programmed to a target resistance of 4 MΩ (representing weight 1) or 10 MΩ (weight 0), followed by a retention test of 1000 s (Supplementary Note 14). As shown in Fig. 5c, the stable retention characteristics ensure reliable synaptic operation. Then the measured IBL,i is used to generate updated reference voltages for next iteration (Supplementary Fig. 14c). During 1000 iterations, the digitized Vo,i of the four p-bits are recorded, with 2 V indicating that the corresponding node is selected and 0 V that it is not (Fig. 5d). Each network state is therefore a four-bit binary number representing a candidate independent set. The histogram in Fig. 5e shows the occurrence probabilities of all observed states calculated over the last 200 iterations and reveals two dominant configurations, “1010” and “0101”, which correspond exactly to the two maximum independent sets of the target graph (Supplementary Fig. 14a). These results confirm that the hardware Boltzmann machine reliably converges to the correct solutions and solves the MIS problem. Many real-world problems, such as RNA secondary structure prediction57, can be reduced to MIS problem and thus solved by FTJ-based Boltzmann machine systems. RNA structure prediction by probabilistic computing RNA is a fundamental functional molecule in living systems and plays a crucial role in gene regulation and targeted drug development. An RNA molecule is made up of four types of nucleotides (A, U, C, and G), arranged in a linear sequence known as its primary structure. Guided by specific nucleotide-pairing rules (mainly A-U, C-G, and sometimes G-U), segments of this sequence fold back to form secondary structures58. Experimental determination of these structures, for example through X-ray crystallography, often requires high-quality crystal samples and is technically challenging, time-consuming, and costly59. As a result, computational approaches for predicting RNA secondary structure have become increasingly important59,60. We focus on the 64-nucleotide human microRNA hsa-miR-375, a crucial molecule involved in insulin secretion regulation and pancreatic cancer suppression, whose primary structure is shown in Fig. 6a (upper panel)61,62. The prediction of its secondary structure can be transformed into an MIS problem by enumerating all possible nucleotide pairings within the RNA sequence (Supplementary Note 15). For hsa-miR-375, the corresponding MIS problem contains 655 nodes. Using the experimental parameters extracted from both types of FTJs, a simulated Boltzmann machine (Fig. 6a, lower panel) comprising 655 p-bits was constructed, with each p-bit representing a valid nucleotide pair. The cycle-to-cycle (C2CVar) and device-to-device (D2DVar) variations of s-FTJs, as well as the read noise and D2DVar of r-FTJs are incorporated into the simulation. Details are provided in the Methods and Supplementary Note 16.Fig. 6RNA secondary structure prediction using an FTJ-based Boltzmann machine.a Unpaired primary structure of hsa-miR-375, along with the corresponding Boltzmann machine employed to predict its secondary structure. b Time traces of four representative p-bits selected from the 655-p-bit Boltzmann machine over the full simulation period. c Evolution of the network energy. d Probability distribution of candidate RNA structures. The red bar denotes the true structure. e Experimentally determined RNA secondary structure of hsa-miR-375. RNA secondary structure prediction using an FTJ-based Boltzmann machine. a Unpaired primary structure of hsa-miR-375, along with the corresponding Boltzmann machine employed to predict its secondary structure. b Time traces of four representative p-bits selected from the 655-p-bit Boltzmann machine over the full simulation period. c Evolution of the network energy. d Probability distribution of candidate RNA structures. The red bar denotes the true structure. e Experimentally determined RNA secondary structure of hsa-miR-375. Figure 6b presents the time evolution of four representative p-bits in the network, where a 2 V (0 V) output indicates the corresponding nucleotide pair is selected (unselected). Each configuration (the collective states of all 655 p-bits) can be mapped to a candidate RNA secondary structure. The trajectory of network energy in Fig. 6c reveals that the network converges after ~8000 iterations, and a total of 10,000 iterations were performed to ensure identification of the lowest energy state. During these iterations, a dynamic-slope annealing scheme with a step schedule was applied, following recent studies63,64. As shown in Fig. 6d, network states sampled from iterations 7000 to 10,000 were used to plot a probability histogram of the candidate secondary structures, along with correlation coefficients evaluating their similarity to the true structure, i.e., the experimentally determined native secondary structure of the hsa-miR-375 (containing 21 nucleotide pairs, as shown in Fig. 6e). Here, a correlation coefficient of 1 indicates perfect consistency. Other commonly used quantitative metrics (Sensitivity, PPV, and F1 score)65 are listed in Supplementary Table 5. The results show that the secondary structure with the highest probability matches the true structure, thus verifying the biological validity of our method. To assess the impact of D2DVar on prediction robustness, we conducted Monte Carlo simulations (Supplementary Fig. 23). The results demonstrate that the parameter variations of our experimental devices are strictly within the safe tolerance region, ensuring reliable performance in large-scale applications. Notably, the probabilistic computing also identifies three metastable configurations, each also comprising 21 nucleotide pairs (Supplementary Fig. 24), which were not observed experimentally but may represent viable structural states. These results demonstrate the practical potential of probabilistic computing in bioinformatics. Furthermore, the proposed framework can be extended to predict more structurally complex tRNAs and mRNAs (Supplementary Note 18), thereby demonstrating its strong generalizability and scalability. Supplementary Table 6 provides the total power, energy, and solving time for experimental 4-node and simulated 655-node systems, and provides a comparison with reported hardware-implemented probabilistic computing systems. This comparison highlights two key advantages of our work: (1) The stochastic neurons (p-bits) and the deterministic synapses are implemented experimentally, culminating in a functional hardware system. This represents a significant step beyond previous studies, which often focused on single-device (p-bit neuron or synapse) functional verification. (2) The energy per operation for p-bits and synapses based on FTJs is orders of magnitude lower than the available data from other technologies. This energy advantage stems from the low current density of the FTJs. In summary, this study proposes and implements a fully hardware-integrated Boltzmann machine based on HZO FTJ devices and circuits. Through controlling the oxygen vacancy concentration in HZO via ALD purge-time engineering, we successfully fabricated s-FTJs with a resistance fluctuation ratio of ~18% and r-FTJs exhibiting reliable resistance switching (RFR ~ 3.7%). The s-FTJ, together with a comparator and resistor, constitutes a low write power p-bit (~76 nW), significantly outperforming existing p-bit circuits. At Vref,i ~ 0.51 V, the p-bit-based TRNG generates stochastic 0/1 sequence validated by the NIST stochasticity tests. Using a four-node MIS problem as a benchmark, the hardware system of a four-p-bit Boltzmann machine converged to the global optimum within 1000 iterations, correctly solving the problem and experimentally validating its capability for combinatorial optimization. Extended simulations to 655 p-bits accurately predicted the secondary structure of a 64-nucleotide RNA, highlighting the promise of ferroelectric probabilistic computing in addressing complex combinatorial optimization challenges. Methods Sample preparation The ~25 nm-thick TiN film, serving as the bottom electrode, was deposited on Si substrate using direct current (DC) magnetron sputtering. Subsequently, the TiO2 interlayer and Hf0.5Zr0.5O2 (HZO) film were grown using atomic layer deposition (ALD) (MNT-PD100Oz) at a temperature of 320 °C, with H2O serving as the oxygen source. Rapid thermal annealing (RTA) was conducted at 600 °C for 20 s in a N2 atmosphere at a pressure of 100 kPa, inducing the o-phase in the HZO film. Following this, top electrodes of platinum (Pt) with a thickness of 30 nm were deposited using DC magnetron sputtering. The device employs a crossbar electrode architecture, with a 10 × 10 µm effective area defined by the overlap between the top and bottom electrodes (Supplementary Note 1). Prior to XPS measurements, the sample underwent an in-situ 30 s etching to remove surface contamination. Structural and electrical characterizations The GIXRD analysis of the HZO/TiO2/TiN heterostructure was conducted using a Rigaku Smartlab 9 kW XRD system, employing Cu Kα 1 radiation (λ = 1.5406 Å). The cross-sectional structure of the Pt/HZO/TiO2/TiN heterostructure was examined with a STEM (Talos F200X) that includes a spherical aberration corrector. PFM phase-voltage and amplitude-voltage loops were obtained with a scanning probe microscope (Asylum Research MFP-3D). Pulsed voltages were generated by a Keithley 3401 pulse generator. The resistance of the FTJ was measured using a Keithley 2450 SourceMeter. Simulation details based on experimental data The simulated p-bit neurons are modeled using the Pi-Vref,i relationship of the s-FTJ (Fig. 4c) and incorporate device non-idealities such as D2DVar and C2CVar. During each computation cycle, the software generates a stochastic output of either 0 or 2 V based on the applied Vref,i. The r-FTJ array is employed to perform synapse operations. In the simulation, the r-FTJ synaptic devices incorporate read variation (σread) and D2DVar to capture their non-ideal characteristics. The synapse matrix is problem-dependent and remains fixed throughout the iterative computation process. Supplementary information Supplementary Information Transparent Peer Review file Supplementary Information Transparent Peer Review file Source data Source Data Source Data