Published benchmarks from peer-reviewed literature — what soft-sensor technology achieves in real industrial settings. Third-party results, each linked to its source.
Chemicals / Polymerization Polymer quality measures are available only from infrequent laboratory analysis, and they depend nonlinearly on reactor temperatures and feed rates.
Nonlinear model predictive control with a first-principles model (material balances with rate expressions, energy balance) controls the laboratory-measured quality in closed loop by setting the reactor temperature profile.
Key finding: In continuous production use since October 2012.
Source: Bindlish (Dow Chemical) — Computers & Chemical Engineering (2015) Pulp & Paper / Tissue Basis weight (g/m²) cannot be measured in real time in the wire section of a tissue machine — only downstream, too late to correct deviations.
Hybrid model: a first-principles model and a 1D convolutional network in parallel; a GRU network weights their outputs according to the current operating regime.
Key finding: RMSE 2.12 g/m² (hybrid) vs 2.68 g/m² (machine learning only) vs 4.72 g/m² (physics only), on 7 months of industrial data.
Source: Othen et al. — RWTH Aachen / Nordic Pulp & Paper Research Journal (2025) Oil & Gas — Downhole Pressure Permanent downhole gauges are expensive and unreliable in harsh well conditions, yet bottom-hole pressure is critical for production optimization and flow assurance.
An LSTM soft sensor trained on wellhead and topside measurements; transfer learning adapts the model to other wells and operating conditions.
Key finding: MAPE consistently below 2 % on real offshore datasets from the Brazilian pre-salt basin.
Source: Fernandes et al. — arXiv (2026) Power Generation — NOx Emissions Real-time NOx prediction is the basis for combustion optimization; working conditions shift with load and fuel, so static models degrade.
Just-in-time learning with a random forest: a local model per prediction, selected by a temporal-spatial similarity index, with moving-window monitoring that triggers updates.
Key finding: R² = 0.93 on the test set; 99.7 % of samples within 15 mg/m³ absolute error — better than all six comparison models.
Source: He & Ding — Sensors / MDPI (2024) Wastewater Treatment COD, TSS and E. coli need sampling and laboratory analysis, so there is no real-time view of treatment performance; online COD analysers are too expensive for onsite systems.
Machine-learning soft sensors on in-line measurements (turbidity, color, pH, NH₄⁺, NO₃⁻, conductivity). Four algorithms compared; SVR was best for COD, cubist regression for TSS.
Key finding: COD: R² = 0.96 (SVR); TSS: R² = 0.99 (cubist) — on 56 data points from a field-tested onsite treatment system; E. coli could not yet be predicted reliably.
Source: Shyu et al. (Univ. South Florida) — ACS Environmental Au (2023) Pharmaceutical / Process Analytical Technology (PAT) Critical quality attributes of tablets — content uniformity, dissolution, hardness — are traditionally measured off-line at the end of the batch, which blocks real-time release testing (RTRT).
Soft sensors infer the quality attributes from in-line PAT instruments (NIR, Raman, laser diffraction) and process data during manufacturing, within the FDA PAT framework.
Key finding: Review of the state of the art of real-time release testing of tablets; the FDA PAT guidance has been in place since 2004.
Source: Markl et al. — International Journal of Pharmaceutics (2020) Semiconductor / Virtual Metrology Metrology steps (film thickness, etch depth, CMP endpoint) are expensive and sampled rarely, leaving quality blind spots between measurements.
Virtual metrology predicts post-process quality from equipment status and process variables, using sampled real metrology to train and correct the models.
Key finding: Review of two decades of virtual-metrology literature: generic steps for a successful VM, the state of the art, and open problems.
Source: Maitra et al. — Expert Systems with Applications (2024) Automotive / Autonomous Vehicles Higher automation levels need more sensors per vehicle — from 8 to 28 in the paper’s overview — adding cost, weight, power and failure points.
Virtual sensors replace or supplement physical units and are validated in full-vehicle simulation; state estimation uses Kalman-family observers and hybrid machine-learning approaches.
Key finding: Virtual-sensor market: USD 1.37 bn (2025) → USD 5.35 bn (2030), as cited from Mordor Intelligence.
Source: Barabás et al. — Sensors / MDPI (2025) 
