In-Situ Sensing-Based Prediction

Updated 10 December 2025

In-situ sensing-based prediction is a real-time method using embedded sensors and machine learning to forecast system dynamics, material properties, and environmental fields.
It leverages multi-modal sensor fusion, Bayesian inference, and deep learning to achieve precise spatiotemporal reconstruction and model adaptation.
Applications range from solar wind and additive manufacturing to soft robotics and quantum metrology, emphasizing uncertainty quantification and closed-loop control.

In-situ sensing-based prediction refers to the direct, real-time acquisition and exploitation of physical measurements obtained from sensors that are co-located with or embedded within the system or process under study, for the purpose of forecasting system state, process outcomes, material properties, or environmental fields. The approach contrasts with retrospective (“ex-situ”) measurements, leveraging immediate access to operational or dynamic system information for timely model adaptation, control, and uncertainty quantification. Modern in-situ prediction workflows often combine physical sensing, machine learning, process modeling, and data fusion, with applications ranging from remote heliospheric observations and environmental field reconstruction to advanced manufacturing and autonomous robotics.

1. Sensing Modalities and Data Integration

In-situ sensing-based prediction encompasses diverse sensor modalities tailored to their target domains:

Physical fields and materials: Embedded thermocouples, capacitive or optical sensors for soft robotics deformation and temperature (Sabelhaus et al., 2021); gas composition sensors (Pirani, mass spectrometry) for chemical vapor environments (Shindler et al., 20 Mar 2025); multiwavelength pyrometers, high-speed infrared, and acoustic sensors for melt pool and defect prediction in laser manufacturing (Chen et al., 21 Apr 2024, Chen et al., 2023).
Remote environments: Spacecraft-borne plasma and magnetic field detectors for solar wind and CME propagation (Rouillard et al., 2017); point oceanographic sensors (temperature, salinity) for marine field mapping (Cutolo et al., 2022).
Networked systems: Wireless sensor node deployments continuous in-situ field monitoring and optimization (Chen et al., 2019).
Autonomous monitoring: Vision, telemetry, and multi-sensor (IMU, GPS) datastreams for UAV-based behavioral monitoring in animal ecology (Kline et al., 23 Jul 2024).
Quantum systems: Electron momentum spectroscopy for strong-field ponderomotive intensity estimation (Maxwell et al., 2020).

Integration of multi-modal data typically relies on precise temporal synchronization and spatial registration, either via explicit control (robotic frameworks) or via mapping to global mesh or volumetric representations (as in digital twins or spatial field fusion). Feature vectors extracted from each modality are concatenated (data-level fusion), transformed and merged (feature-level fusion), or post-processed with ensemble or Bayesian strategies (decision-level and Bayesian fusion) (Chen et al., 21 Apr 2024).

2. Mathematical and Statistical Prediction Frameworks

Underlying in-situ prediction architectures exhibit broad methodological diversity but consistently exhibit a combination of sensing, feature extraction, and statistical or machine learning-based mapping to state, process, or field predictions.

Spatiotemporal Field Reconstruction: Bayesian hierarchical models—for instance, integrating spatially misaligned point (in-situ) and areal (satellite) data—can be realized via SPDE-based GMRFs with INLA, incorporating projections from mesh nodes to both sensor and remote observation locations, AR(1) time dynamics, and Matérn spatial covariance (He et al., 9 Jan 2024).
Optimal Sensor Selection and Nonlinear Field Forecasting: Sensor placement is optimized via a determinant-maximization over PCA mode projections (QR pivoted DEIM) to maximize information content of in-situ samples. Deep learning architectures (compression-LSTM-MLP) reconstruct and forecast full fields from these measurements (Chen et al., 2019).
Physics-based Propagation and Data Assimilation: For heliospheric structures, mathematical kinematic models (ballistic, drag-based, Archimedean spiral, Parker spiral length) provide time-resolved predictions of transient evolution based on both remote fits and in-situ initialization; physical equations for CME/CIR and SEP transport are integrated and dynamically updated using in-situ measured parameters (Rouillard et al., 2017).
Machine Learning and Neural Models: In manufacturing, feed-forward neural networks (ANN) and Bayesian neural networks (BNN) directly relate process parameters, sensor data, or even RGB images (digital image colorimetry) to outputs such as etch depth, with explicit uncertainty quantification from MC Dropout (Kang et al., 3 May 2025). In soft robotics, LSTM networks trained on temperature and deflection sensor data capture hysteresis-dominated actuator behavior, achieving near-sensor-limited prediction performance (Sabelhaus et al., 2021).
Quantum Estimation Theory: Quantum and classical Fisher information for in-situ strong-field ionization are derived to set quantum-limited bounds on laser intensity measurement, rigorously quantifying measurement uncertainty as functions of pulse parameters and electron detection fidelity (Maxwell et al., 2020).

3. Sensor Fusion and Deep Learning Approaches

Advanced in-situ prediction frameworks increasingly leverage sensor fusion and deep learning:

Digital Twins and Multimodal Quality Prediction: In additive manufacturing, spatially and temporally fused datasets from synchronized acoustic, thermal, vision, and 3D point cloud acquisition are used to construct high-dimensional feature sets. Machine learning (gradient-boosted trees, NNs, planned spatio-temporal CNN/RNN hybrids) then map sensor signatures to voxel-resolved quality or defect metrics, enabling real-time toolpath correction and remediation (Chen et al., 2023, Chen et al., 21 Apr 2024).
Fused Remote and In-situ Ocean State Estimation: CLOINet combines classical Gaussian optimal interpolation (OI) with self-supervised cluster-based feature extraction from remote images (e.g., SST, SSH), yielding cluster-informed covariances and enhancing fine-scale resolution by up to 50% over OI (Cutolo et al., 2022). Multiple neural modules (OINet, CluNet, RefiNet) are trained end-to-end, handling both simulated and real-world data fusion scenarios.
Imageomics and Autonomously-Guided Data Collection: Autonomous UAV systems integrate image-based animal localization and behavior inference (YOLO + temporal CNNs), with real-time telemetry-guided navigation policies (decision-theoretic, imitation-learned) to maximize collection of high-utility in-situ video for trait inference with optimal flight-path control (Kline et al., 23 Jul 2024).
Federated and Collaborative Edge Learning: In cellular networks, distributed LSTM models distinguish between global prediction layers and device-specific hardware/stack “sensitivity adapters,” trained via federated aggregation of in-situ data from diverse edge devices for robust, generalizable throughput prediction (Sen et al., 2023).

4. Performance Metrics, Uncertainty Quantification, and Practical Constraints

Evaluation and deployment of in-situ sensing-based predictors require carefully chosen metrics and robust uncertainty quantification:

Quantitative metrics: RMSE, explained variance, Pearson correlation, effective resolution (from NSR), F1, ROC AUC, and mean square error (on both field prediction and process outputs), as well as Cramér–Rao lower bounds for quantum-limited metrology.
Uncertainty quantification: Bayesian neural inference (BNN, MC Dropout), coverage analysis, and analytically derived Fisher information bounds are essential in safety-critical or high-precision applications (Kang et al., 3 May 2025, Maxwell et al., 2020).
Implementation practicalities: Real-time sensor fusion and model inference are achieved via alignment/synchronization (RTOS, timestamped pipelines), data normalization, robust denoising, and compact architectures amenable to edge/gpu deployment.
Limiting factors: Data sparsity, sensor calibration, cross-modal time alignment, domain shift, and class imbalance must be addressed via model regularization, augmentation, and careful experimental design.

5. Case Studies and Domain-specific Impact

In-situ sensing-based prediction enables domain-specific advances:

Domain	Key Approach/Metric	Representative Impact
Solar wind & CME	Kinematic models & in-situ shock timing	CME arrival predictions within 4–8 h at 1 AU, SEP onsets within tens of minutes (Rouillard et al., 2017)
Additive manufacturing	Multi-sensor fusion + ML defect classification	AE+optical fusion achieves ~99% accuracy, with closed-loop defect correction in robotic LDED (Chen et al., 2023, Chen et al., 21 Apr 2024)
Oceanography	Deep learning cluster-informed OI	40% RMSE reduction, 50% finer resolved scales for deep salinity reconstructions (Cutolo et al., 2022)
Semiconductor etch	ANN/BNN + DIC	Non-contact in-situ RGB-based ANN reaches 10.38 nm² MSE, with reliable coverage for 98% of validation (Kang et al., 3 May 2025)
Soft robotics	LSTM on temperature/angle sensors	10+ min open-loop deflection prediction with RMSE matching sensor accuracy (Sabelhaus et al., 2021)
Quantum metrology	QFI, CFI from in-situ momentum spectra	25× uncertainty reduction over yield-based intensity inference; minimum uncertainty ~0.003% (Maxwell et al., 2020)

These case studies demonstrate substantial improvements over ex-situ or single-sensor baselines, real-time responsiveness, and enhanced physical insight.

6. Open Challenges and Research Directions

Outstanding challenges include:

Standardization and reproducibility: Protocols for sensor fusion, data alignment, ML benchmarking, and open datasets are needed, particularly for production-scale manufacturing (Chen et al., 21 Apr 2024).
Uncertainty propagation and decision-making: Robust calibration, process-aware uncertainty models, and adaptive control logic for closed-loop operation remain developing research areas.
Sensor optimization and field redeployment: Active learning for sensor placement (Chen et al., 2019), transfer learning across devices/processes, and development of adaptive, self-tuning controllers.
Multiscale and hierarchical modeling: Integration of real-time (ms), layer/region (s–min), and macro-scale (geometry) data for defect prediction and adaptive quality remediation.
Fusion of physics and data-driven models: Hybrid architectures that combine mechanistic process equations, clustering, and deep learning for robust, physically interpretable prediction (Cutolo et al., 2022, Maxwell et al., 2020).

A plausible implication is that as in-situ prediction frameworks mature, their unification of distributed sensing, statistical inference, physical modeling, and real-time actuation will underpin advances toward truly self-adaptive, zero-defect, and quantum-limited systems across science and engineering domains.