BLISSNet: Deep Operator Learning for Fast and Accurate Flow Reconstruction from Sparse Sensor Measurements
| Authors | Maksym Veremchuk et al. |
| Year | 2026 |
| Field | Machine Learning |
| arXiv | 2602.24228 |
| Download | |
| Categories | cs.LG |
Abstract
Reconstructing fluid flows from sparse sensor measurements is a fundamental challenge in science and engineering. Widely separated measurements and complex, multiscale dynamics make accurate recovery of fine-scale structures difficult. In addition, existing methods face a persistent tradeoff: high-accuracy models are often computationally expensive, whereas faster approaches typically compromise fidelity. In this work, we introduce BLISSNet, a model that strikes a strong balance between reconstruction accuracy and computational efficiency for both flow reconstruction and nudging-based data assimilation. The model follows a DeepONet-like architecture, enabling zero-shot inference on domains of arbitrary size. After the first model call on a given domain, certain network components can be precomputed, leading to low inference cost for subsequent evaluations on large domains. Consequently, the model can achieve faster inference than classical interpolation methods such as radial basis function or bicubic interpolation. This combination of high accuracy, low cost, and zero-shot generalization makes BLISSNet well-suited for large-scale real-time flow reconstruction and data assimilation tasks.
Engineering Breakdown
Plain English
BLISSNet solves the problem of reconstructing fluid flow fields from sparse sensor measurements—a critical challenge in engineering and scientific computing where sensors are far apart and fluid dynamics involve complex, multiscale phenomena. The paper introduces a deep operator learning model that achieves a significant balance between reconstruction accuracy and computational speed, addressing a long-standing tradeoff where accurate models are slow and fast models lose fidelity. The architecture follows a DeepONet-like design, meaning it can work on fluid domains of arbitrary size without retraining (zero-shot inference), and certain network components can be cached after the first call to accelerate subsequent inference on the same domain. This approach also works for nudging-based data assimilation, a technique used in weather forecasting and other domains where real-time measurements are incrementally incorporated into simulations.
Core Technical Contribution
The core novelty is a DeepONet-inspired operator learning framework specifically engineered to handle the sparse-observation, fast-inference regime in fluid reconstruction. Unlike standard neural networks that map fixed input sizes to fixed output sizes, operator networks learn maps from functions (sensor measurements) to other functions (full flow fields), enabling generalization across domain sizes and geometries. BLISSNet's key innovation appears to be the introduction of component caching—after processing a domain once, certain network branches can be cached and reused, dramatically reducing computational cost on repeated queries without sacrificing accuracy. This design simultaneously solves two hard problems: achieving near-physics-based fidelity while maintaining real-time inference speeds suitable for online data assimilation and monitoring systems.
How It Works
BLISSNet takes sparse sensor measurements (temperature, velocity, pressure at scattered locations) as input and passes them through an encoder network that learns a latent representation of the flow state. This representation feeds into a DeepONet-style operator that learns the mapping from the sensor latent space to the full spatial flow field across the entire domain. The network outputs velocity, pressure, and other flow quantities at grid points covering the entire domain, even where no sensors exist. The key efficiency mechanism is architectural: certain network components (likely the basis functions in the DeepONet terminology) are computed once per domain and cached as tensors; subsequent queries only require fresh computation of lightweight decision/adaptation layers that condition on new sensor inputs. This two-stage design—expensive one-time setup followed by cheap online updates—mirrors how numerical weather prediction systems work, where expensive preprocessing is amortized over many forecast cycles.
Production Impact
For engineers building real-time fluid monitoring systems (industrial manufacturing, HVAC control, aerodynamic optimization), BLISSNet eliminates the speed-accuracy tradeoff that currently forces a choice between expensive CFD solvers and crude interpolation methods. A production system could deploy this model to consume sparse IoT sensor streams and output full-field reconstructions in milliseconds, enabling closed-loop control that previously required either running expensive simulations in parallel or accepting poor spatial resolution. The zero-shot generalization to arbitrary domain sizes means you can train once on representative geometries and directly apply the model to new facility layouts without retraining—a major practical advantage in industrial settings with many similar but not-identical installations. The caching mechanism also reduces infrastructure requirements: after the initial setup phase, inference becomes lightweight enough to run on edge devices or embedded systems, avoiding constant cloud communication. However, you will need high-quality labeled training data (full ground-truth flow fields from simulations or experiments) and must validate that the model generalizes to flow regimes and sensor configurations not seen during training.
Limitations and When Not to Use This
BLISSNet assumes the sparse sensor data faithfully represent the underlying flow and doesn't explicitly handle sensor noise, calibration drift, or missing sensors—these are common failure modes in real deployments where sensor health degrades over time. The paper doesn't discuss what happens when the true flow regime shifts outside the training distribution (e.g., transition from laminar to turbulent flow, or operation at Reynolds numbers far from training conditions); operator networks can extrapolate worse than physics-informed models in these regimes. Scalability to very high-dimensional domains (3D flows with fine temporal resolution) is unclear—the token budget limitation in the abstract suggests the full paper may address this, but operator networks typically face curse-of-dimensionality challenges that get worse with problem size. The approach also requires paired training data (sparse measurements + ground truth), which can be expensive to generate for complex real-world flows; for domains where only sparse data exist naturally (like oceanography), you may need to run expensive simulations to create training sets, reducing the practical advantage over traditional surrogate models.
Research Context
BLISSNet builds directly on DeepONet (Chen & Perdikaris, 2019-2021), which pioneered the use of operator networks for learning mappings between infinite-dimensional function spaces, but optimizes that framework specifically for the sparse-observation, zero-shot-generalization regime. The work is positioned against both classical reduced-order models (POD, autoencoders) and recent neural surrogate approaches (FNO, GraphNeuralOperator); classical methods are interpretable but slow, while FNO variants are fast but require retraining for new domains or solving new PDEs. The contribution fills a gap in the literature between two communities: the scientific computing crowd doing data assimilation and inverse problems, and the deep learning community focused on operator learning, by making operator networks practical for real-time applications with hard computational budgets. This work likely opens a research direction in hardware-aware operator learning where caching and component factorization become central design principles, similar to how LoRA and other parameter-efficient finetuning methods revolutionized large language model deployment.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::