QLIF-CAST: Quantum Leaky-Integrate-and-Fire for Time-Series Weather Forecasting
Alberto Marchisio 1 2 , Aayan Ebrahim 1 2 , Nouhaila Innan 1 2 , Muhammad Kashif 1 2 , Muhammad Shafique 1 2 · 2026
Read QLIF-CAST: Quantum Leaky-Integrate-and-Fire for Time-Series Weather Forecasting by Alberto Marchisio 1 2 , Aayan Ebrahim 1 2 , Nouhaila Innan 1 2 , Muhammad Kashif 1 2 , Muhammad Shafique 1 2 (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-18333.txt.
Introduction
Alberto Marchisio's 2026 paper adapts quantum leaky integrate-and-fire neurons to short-term multivariate weather forecasting. QLIF-CAST encodes neuron excitation as single-qubit superpositions, driven by rotation gates and relaxation decay, in a hybrid recurrent design. Against a matched classical baseline, it cuts mean squared error by 15.4 percent and mean absolute error by 4.4 percent. Against quantum LSTM and quantum network models on air quality and wind benchmarks, it converges in up to 94 percent less training time. Hardware runs on IBM Marrakesh deviate only 1.2 percent from simulation.
Catalog record
| CREATOR | Alberto Marchisio 1 2 , Aayan Ebrahim 1 2 , Nouhaila Innan 1 2 , Muhammad Kashif 1 2 , Muhammad Shafique 1 2 |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 42495 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.18333
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
QLIF-CAST: Quantum Leaky-Integrate-and-Fire for Time-Series Weather Forecasting Alberto Marchisio 1 2 , Aayan Ebrahim 1 2 , Nouhaila Innan 1 2 , Muhammad Kashif 1 2 , Muhammad Shafique 1 2 Abstract Accurate and efficient time-series forecasting remains a challenging problem for both classical and quantum neural architectures, particularly in multivariate environmental settings. This work adapts the Quantum Leaky Integrate-and-Fire (QLIF) spiking neural network for time-series regression tasks, specifically short-term multivariate weather forecasting. We extend QLIF beyond classification and demonstrate its applicability to continuous-valued prediction problems. The QLIF-CAST model encodes neuron excitation states as single-qubit quantum superpositions, driven by R x R_{x} rotation gates and T1 relaxation decay, and is embedded within a hybrid quantum-classical recurrent architecture. We conduct two distinct evaluations. First, a controlled comparison against a parameter-matched classical LIF baseline on a multivariate weather dataset shows that QLIF-CAST achieves 15.4% lower MSE and 4.4% lower MAE, demonstrating that quantum neuronal dynamics reduce prediction error over classical equivalents. Second, a cross-domain comparative analysis with state-of-the-art quantum LSTM (QLSTM) and quantum neural network (QNN) models on air quality and wind speed benchmarks reveals that QLIF-CAST converges in up to 94% less training time, occupying a distinct position in the speed-error trade-off space. Hardware verification on IBM Marrakesh (156-qubit QPU) confirms reliable circuit execution with only 1.2% average deviation from simulation. I Introduction Short-term multivariate weather forecasting requires models capable of capturing complex temporal dependencies across correlated physical variables [ 1 , 2 ] . Traditional recurrent architectures such as LSTM often trade computational cost for predictive quality [ 3 , 4 ] , while classical spiking neural networks offer event-driven efficiency but lack expressive state representations [ 5 , 6 ] . Quantum spiking models combine the discrete, event-driven nature of spikes with quantum superposition and interference, potentially offering richer state representations than classical neurons within the same parameter budget [ 7 , 8 ] . However, while the QLIF model [ 9 , 10 ] has demonstrated strong results on image classification (MNIST, 98.1%), its application to time-series regression remains unexplored. In this paper we address this gap through two complementary evaluations: 1. Controlled quantum vs. classical comparison: QLIF-CAST is compared against a parameter-matched classical LIF baseline, identical in all respects except the neuronal update rule. This isolates the effect of quantum dynamics on prediction error for multivariate weather forecasting. 2. Comparative analysis against published quantum models: QLIF-CAST is evaluated on benchmark tasks used by QLSTM [ 11 ] and LSTM-QNN [ 12 ] , characterizing the speed-error trade-off between simple single-qubit circuits (in QLIF-CAST) and deeper variational quantum circuits (in QLSTM and LSTM-QNN). Contributions: This work makes the following contributions: 1. Novel application domain for QLIF: We extend the QLIF spiking neuron, originally demonstrated on image classification, to continuous-valued multivariate time-series regression, establishing it as a viable architecture for temporal forecasting tasks. 2. Systematic speed-error characterization: We provide the first systematic comparison of QLIF-CAST’s depth-2, classically-driven single-qubit circuits (with no trainable quantum parameters) against deeper variational quantum architectures (QLSTM, LSTM-QNN) across multiple real-world forecasting benchmarks, quantifying the trade-off between training efficiency and prediction accuracy. 3. Cross-domain benchmark evaluation: QLIF-CAST is evaluated across three distinct environmental forecasting domains (meteorological, air-quality, and wind-energy) without modifying the core quantum circuit, demonstrating architectural generality. 4. Practical vectorized implementation: We introduce a batched, vectorized circuit execution strategy in which all rotation angles across all neurons and timesteps in a batch are executed in a single parallel circuit call, yielding a 500{\times} speedup over sequential execution and making QLIF-CAST training on real-scale datasets feasible under classical simulation. 5. Near-term QPU hardware validation: We provide empirical validation of the QLIF-CAST circuit on a real 156-qubit QPU (IBM Marrakesh), contributing a reproducible hardware benchmark for shallow, classically-driven quantum circuits in the QML literature. II Background and Related Work II-A Spiking Neural Networks for Time-Series Spiking neural networks (SNNs) process information through discrete spike events rather than continuous activations, offering a computational model closer to event-driven neural dynamics than conventional artificial neural networks [ 13 ] . Among the most widely used spiking neuron models, the Leaky Integrate-and-Fire (LIF) neuron combines input integration, membrane leakage, and threshold-triggered spike emission in a simple dynamical system [ 14 , 15 ] . Because of this balance between simplicity and temporal modeling capability, LIF-based architectures have become a common starting point for sequence processing in spiking systems. Prior work has shown that SNNs can be effective for temporal tasks such as speech recognition and event-driven sensor processing [ 16 , 17 ] . However, most existing studies emphasize classification or detection, while comparatively fewer investigate continuous-valued forecasting. In particular, the use of SNNs for multivariate regression remains limited, leaving open the question of whether spike-based dynamics can support accurate prediction in practical time-series settings. II-B Quantum Neural Networks Quantum neural networks aim to exploit quantum-mechanical effects such as superposition and entanglement to construct models with richer internal representations than their classical counterparts [ 18 ] . A common approach is to use variational quantum circuits (VQCs), in which parameterized quantum gates are optimized with classical gradient-based routines [ 19 ] . Such hybrid quantum-classical models have shown promise in the noisy intermediate-scale quantum (NISQ) setting, where near-term quantum devices are available but remain constrained by limited depth and noise sensitivity [ 20 ] . However, VQC-based models face practical limitations. As circuit depth grows, gate noise accumulates and the cost of training increases, particularly when gradients are estimated through parameter-shift rules. In recurrent settings, where quantum subcircuits may be invoked repeatedly across time steps, this overhead can become a substantial bottleneck. These considerations motivate interest in quantum architectures that retain useful quantum dynamics while avoiding the training cost of deeply parameterized circuits. II-C QLIF Neuron Model The Quantum Leaky Integrate-and-Fire (QLIF) model [ 10 ] adapts the logic of spiking neurons to a single-qubit setting by representing neuronal excitation as the probability of measuring the qubit in the state |1\rangle . At each time step, the excitation state is updated through R_{x} rotations that encode the current state and the incoming signal, yielding a compact quantum version of state accumulation and decay. Unlike standard variational quantum models, the circuit does not introduce separately trainable quantum gate parameters. Instead, the rotation angles are computed directly from classical network outputs (see Fig. 1 ). This design choice is important for two reasons. First, it preserves a simple quantum circuit structure that is easier to execute and analyze than deeper variational alternatives. Second, it avoids the additional optimization burden associated with trainable quantum parameters, making QLIF closer in training complexity to a classical recurrent component than to a fully variational quantum layer. Prior work has demonstrated the effectiveness of QLIF in image classification, which motivates examining whether the same neuron model can support continuous-valued temporal prediction. Fig. 1: QLIF neuron model. The qubit is initialized in |0\rangle , rotated by R_{x}(\phi) to encode the current excitation state, then rotated by R_{x}(\theta_{\text{input}}) to integrate the input signal. Measuring the qubit yields the updated excitation probability \alpha_{\text{new}}=\sin^{2}(\frac{\phi+\theta_{\text{input}}}{2}) . II-D Classical Leaky Integrate-and-Fire The Classical LIF neuron [ 15 ] , shown in Fig. 2 , models membrane potential dynamics via an exponential decay rule and