Real-time and near-real-time process variables
BIOMEDANA SOFT SENSORS · PREDICTIVE PROCESS INTELLIGENCE
Predict what you cannot continuously measure. Understand what drives it.
Combine biology, physics, mechanistic models, process data and machine learning to estimate hard-to-measure variables and surface emerging quality risks earlier.
From upstream cell culture and fermentation to downstream purification—predictions stay connected to evidence, uncertainty and process context.
SOFT SENSOR · FED-BATCH 24-071
MODEL v2.4 · ILLUSTRATIVEESTIMATED NOW4.18 g/L
PREDICTION INTERVAL±0.24
MODEL STATUSWithin domain
PredictedOffline assayInterval
Day 0Day 3Day 7Day 10Day 14
INFLUENTIAL INPUTSVCDFeedDOpHTemperature
Illustrative interface and data. A deployed model requires process-specific data, qualification and defined operating limits.
01 — EARLIER PROCESS VISIBILITY
Between the assay points,
the process keeps moving.
Soft sensors complement physical sensors and laboratory analytics. They make useful estimates available earlier and more frequently—without hiding the assumptions or limitations behind them.
Titer, metabolites, biomass and selected quality attributes
Process drift and emerging quality risk earlier
Influential inputs, uncertainty and model applicability
Historians, PAT, laboratory data and digital twins
Model versions, review, monitoring and change control
02 — WHAT IS A SOFT SENSOR?
A virtual measurement built from signals, science and context.
A soft sensor uses mathematical, statistical, mechanistic or machine-learning models to estimate a variable that cannot be measured directly, continuously or economically with conventional instrumentation.
A bioreactor may continuously report dissolved oxygen, pH, temperature, agitation and gas flow while viable cell density, metabolites and product titer arrive only periodically. A qualified model can estimate those variables between laboratory measurements.
Inline process signals+Offline reference assays+Scientific relationships→Explainable estimate
with uncertainty
with uncertainty
03 — HOW IT WORKS
From process data to
scientifically informed predictions.
Every estimate remains connected to its inputs, model version, process context and intended use.
Connect process data
Bring together bioreactors, equipment, historians, PAT instruments, laboratory systems and manufacturing records.
Contextualize the process
Align signals by batch, phase, material, equipment, sampling time and operating conditions.
Configure the model
Select statistical, mechanistic, machine-learning or hybrid methods for the intended use.
Generate predictions
Update estimates as suitable inputs arrive, with deployment-specific frequency and latency.
Explain and monitor
Evaluate performance, uncertainty, influential features, drift and recalibration needs.
04 — UPSTREAM + DOWNSTREAM
One soft-sensing foundation.
Different process questions.
See growth, metabolism and productivity sooner.
- Viable cell density and biomass
- Glucose, lactate, glutamine and ammonia
- Oxygen uptake and metabolic state
- Product titer and harvest outlook
- Process-specific quality-risk indicators
EXAMPLEPredict mAb titer before the next offline assay.
Combine cell density, feed history, DO, pH, temperature, metabolites and historical titer to track evolving productivity and inform harvest planning.
Anticipate separation performance and recovery risk.
- Chromatography breakthrough and loading
- Recovery and elution characteristics
- Membrane fouling and flux decline
- Impurity and quality indicators
- Intermediate and cumulative yield
EXAMPLEEstimate Protein A breakthrough before it occurs.
Combine loading history, UV, flow, resin characteristics and binding relationships to evaluate strategies and surface risk earlier.
05 — HYBRID ARCHITECTURE
Use the model the process needs.
Not the most fashionable one.
Bioprocess data can be limited and conditions change across materials, equipment and scale. BioMedAna supports multiple model families so scientific knowledge can constrain what data alone cannot.
Mechanistic
Mass balances, kinetics, substrate consumption, oxygen transfer and transport phenomena.
Statistical
PCA, PLS regression, chemometrics and multivariate process models.
Machine learning
Tree ensembles, Gaussian processes, support vector regression and neural networks.
Dynamic
State-space models, Kalman filtering and time-series estimation.
Hybrid
Scientific constraints and mechanistic states combined with data-driven correction.
Uncertainty-aware
Intervals, sensitivity, applicability-domain and out-of-distribution indicators where supported.
06 — APPLICATION CATALOG
Candidate applications across
process, performance and quality.
Each application needs suitable inputs, reference measurements, development and validation. Availability depends on the process, modality, instrumentation and deployment scope.
Cell growth & biomass
Viable and total cell density, viability, biomass, specific growth rate
Process-specific qualification requiredMetabolites & nutrients
Glucose, lactate, glutamine, ammonia and consumption rates
Process-specific qualification requiredBioreactor performance
OUR, kLa, CER, respiratory quotient and metabolic state
Process-specific qualification requiredProductivity
Product titer, specific productivity and predicted harvest titer
Process-specific qualification required07 — CONNECTED TO THE BIOMEDANA PLATFORM
From an isolated estimate to
connected bioprocess intelligence.
01
Bio Hub
Connect and contextualize process measurements, assays, equipment and manufacturing records with lineage intact.
Explore Bio Hub ↗02Bio Twin
Combine soft-sensor estimates with mechanistic and hybrid simulation for scenario evaluation.
Explore Bio Twin ↗03Bio Agents
Compare batches, investigate drivers and prepare source-backed monitoring summaries for review.
Explore Bio Agents ↗04Bio OS
Control versions, access, approvals, audit trails and deployment through the model lifecycle.
Explore Bio OS ↗08 — VALIDATION + LIFECYCLE CONTROL
Accuracy matters.
Credibility and control matter too.
For regulated biomanufacturing, model performance must be evaluated against its intended use, limitations and impact on decisions.
Define
Intended use, inputs, reference methods, assumptions and acceptance criteria.
Evaluate
Held-out or prospective data, relevant operating conditions and data-leakage controls.
Qualify
Prediction error, uncertainty, sensitivity and performance outside the calibration domain.
Monitor
Residuals, input quality, process drift and triggers for investigation or recalibration.
Control
Model versions, training data, configuration, approvals and change records.
BioMedAna’s governance approach is designed to support applicable expectations under 21 CFR Part 11, GAMP 5 and relevant guidance for AI-enabled systems. Actual compliance and validation requirements depend on intended use, deployment and organizational procedures.
09 — DEPLOYMENT
Start with one measurable challenge.
Expand from demonstrated value.
Use case
Target variable, decision and data readiness
Develop
Select and configure the model approach
Evaluate
Compare against independent references
Integrate
Connect qualified outputs to workflows
Monitor
Track performance and manage change
Typical data sourcesProcess historiansSCADA / DCS / MESRaman / NIR / PATLIMS and assaysBatch eventsEquipment and material metadata
Common questions about soft sensors
What is the difference between a physical sensor and a soft sensor?
A physical sensor directly measures a property through instrumentation. A soft sensor estimates a variable from other measurements and a mathematical or computational model.
Can soft sensors predict product quality attributes?
Selected attributes can be modeled when representative process data, analytical measurements and validated predictive relationships are available. Feasibility varies by attribute and process.
Do soft sensors require large datasets?
Not always. Mechanistic, statistical and hybrid models can help in data-constrained settings, but the data must still be suitable and representative for the intended use.
Can soft sensors work in real time?
Yes, when input data arrives with suitable frequency and latency and the model executes within the required timeframe. Some applications are better suited to near-real-time or phase-based predictions.
Can soft sensors be used in GMP manufacturing?
Potentially, subject to intended-use assessment, qualification or validation, data integrity, lifecycle controls and applicable quality-system requirements. They do not automatically replace an approved analytical method.
How are soft sensors different from digital twins?
A soft sensor estimates an unmeasured or infrequently measured variable. A digital twin represents selected aspects of a process for simulation, prediction and scenario analysis. They can work together.
START WITH ONE HIGH-VALUE SIGNAL
Turn process data into predictive bioprocess intelligence.
Bring us a target variable, the available process signals and the decision you want to improve.