SIMULATION-DRIVEN DIGITAL TWIN

A living model
that learns from
the machine it
mirrors.

Physics-based simulation calibrated by real-time IoT telemetry. Not a static replica — a computational companion that evolves with its physical twin, predicts its future, and closes the loop between field reality and engineering assumptions.

DIGITAL TWIN STATUS — TURBINE UNIT GT-7200-A · 18,400 OPERATING HOURS
PHYSICAL TWIN (IoT TELEMETRY)
Bearing temp
142°C
Vibration RMS
4.2 mm/s
Rotor speed
12,840 RPM
Exhaust temp
628°C
DIGITAL TWIN (PHYSICS MODEL)
Predicted temp
138°C
Predicted vib
3.1 mm/s
Model speed
12,840 RPM
Predicted exh
624°C
Drift detected: vibration divergence +1.1 mm/s — physical exceeds model prediction by 35%
Root cause hypothesis: bearing degradation consistent with 18,400h operating profile. RUL estimate: 2,200 hours remaining. Maintenance window recommended within 45 days. Trace engine linked — as-built bearing lot TN-2023-1804 flagged for batch correlation.
CALIBRATION DRIFT — MAINTENANCE ADVISORY
THE SIMULATION GAP

Simulations predict how a product should behave. Digital twins reveal how it actually does.

The gap between design-time simulation and field-time reality is where failures hide, maintenance windows are missed, and engineering assumptions go unchallenged.

$50B
Annual cost of unplanned equipment downtime in manufacturing globally
ORAA RESEARCH 2025
35%
Reduction in latency achieved by edge-deployed digital twins vs. cloud-only architectures
NATURE SCIENTIFIC REPORTS 2025
Hybrid
Physics + data fusion models outperform either approach alone for accuracy and interpretability
SCIENCEDIRECT 2026
30%
Typical reduction in prototyping costs when digital twins replace physical test campaigns
NEXUS DEPLOYMENT DATA

A simulation is a prediction made once, under assumed conditions, at a point in time. A digital twin is a prediction that is continuously recalibrated against reality. The simulation says the bearing temperature should be 138°C. The IoT sensor says it is 142°C. The simulation says vibration should be 3.1 mm/s. The accelerometer says it is 4.2 mm/s. Mirror does not average the difference. It asks: why does reality deviate from the physics model?

The answer — bearing degradation at 18,400 hours, consistent with the operating load profile from the IoT telemetry stream — becomes a maintenance prediction grounded in physics, not statistics. The digital twin does not replace the simulation. It extends the simulation into the operating life of the product, calibrating the physics model with every data point the field generates. When the twin’s prediction diverges from reality, the divergence itself is diagnostic: it tells you what is changing about the physical system that the original model did not anticipate.

WHY MIRROR

Five capabilities that extend simulation from design time into operating life.

Physics-Data Hybrid Architecture
Combines first-principles physics models (FEA, CFD, thermal) with machine learning trained on operational sensor data. The hybrid delivers higher accuracy than either approach alone, with uncertainty quantification at every prediction point.
Physics interpretability + data adaptability in a single model
Sensor-to-Simulation Mapping
Maps every IoT sensor location to the corresponding node in the physics model. Enables point-by-point comparison between what the simulation predicts and what the sensor observes — at matched spatial locations and matched operating conditions.
Physical sensor ↔ virtual model node pairing for real-time comparison
Remaining Useful Life Prediction
Predicts time-to-failure based on the actual operating loads experienced by this specific unit — not fleet averages. Physics-based fatigue curves from Nexus simulation, calibrated by the real load history from Echo IoT telemetry, produce unit-specific RUL estimates.
RUL based on physics + actual loads, not statistical fleet averages
Drift Detection & Anomaly Diagnosis
When the physical twin diverges from the digital twin’s prediction, the divergence pattern is itself diagnostic. Temperature drift suggests thermal degradation. Vibration drift suggests mechanical wear. Correlated multi-parameter drift suggests systemic aging.
Divergence patterns diagnose degradation mode before failure
Closed-Loop Design Feedback
When field data persistently deviates from simulation predictions across the fleet, the deviation feeds back into the design process via Cascade as a corrective ECR. The simulation model that created the twin is updated — closing the loop between field reality and engineering assumptions.
Field reality updates design assumptions via Cascade ECR pipeline
DIGITAL TWIN INTELLIGENCE ENGINES

Eight engines. Every asset mirrored.

From physics-informed neural networks to fleet-wide twin orchestration — Mirror operates eight engines that transform static simulations into living computational companions.

01
Physics-Data Hybrid Architecture
First-principles foundation · ML correction layers · Uncertainty quantification · PINNs integration
▼
Purely physics-based models capture the fundamental behavior but cannot account for manufacturing variability, material aging, and operating condition nuances that accumulate over thousands of hours. Purely data-driven models can fit observed behavior but lack physical interpretability and fail unpredictably outside their training distribution. Mirror’s hybrid architecture uses the physics model as the structural backbone and trains ML correction layers on the residual between physics prediction and observed reality. The physics model provides the base prediction grounded in first principles (conservation laws, constitutive equations, boundary conditions from Nexus). The ML layer learns the systematic bias — the consistent difference between what physics predicts and what the field observes — which typically represents manufacturing variability, material aging effects, and environmental factors not captured in the original simulation.
Physics-Informed Neural Networks (PINNs): Neural networks constrained by physical laws embedded as loss function terms. The network cannot learn solutions that violate conservation of energy, momentum, or mass — ensuring predictions remain physically plausible even in data-sparse regions of the operating envelope
Reduced-order model foundation: The physics backbone uses ROMs generated by Nexus’s Reduced-Order Model Factory — not the full-fidelity FEA/CFD model. ROMs execute in milliseconds (vs. hours for full models), enabling real-time twin updates at sensor sampling rates up to 1 kHz
Uncertainty quantification: Every prediction from the hybrid model includes a confidence interval derived from both the physics model’s parametric uncertainty and the ML layer’s epistemic uncertainty. Wide confidence intervals signal regions where the twin needs more calibration data
Hybrid
Physics backbone + ML correction
PINN
Physics-constrained neural networks
1 kHz
Real-time update rate (ROM backbone)
UQ
Uncertainty quantification per prediction
02
Sensor-to-Simulation Mapping
Spatial registration · Sensor-model node pairing · Multi-physics alignment · Data quality gating
▼
A digital twin is only as valid as the comparison between its predictions and reality. That comparison requires precise spatial registration: which physical sensor corresponds to which node in the simulation mesh? Mirror’s sensor mapping engine establishes a persistent link between each IoT sensor (identified by sensor ID, installation location, measurement axis) and its corresponding computational node in the Nexus simulation model. This mapping is maintained through engineering changes — when Cascade processes an ECO that moves a sensor location or modifies the simulation mesh, the mapping is automatically updated.
Multi-physics alignment: Maps thermal sensors to thermal model nodes, vibration sensors to structural dynamics model nodes, pressure sensors to CFD model nodes, and strain gauges to stress analysis nodes — across different physics domains within the same twin
Data quality gating: Sensor data passes through quality filters before comparison: out-of-range rejection, frozen-signal detection, noise floor validation, and sampling rate verification. Only validated data is used for model calibration — preventing sensor failures from corrupting the twin
Virtual sensor inference: Where physical sensors cannot be installed (internal cavities, rotating components, high-temperature zones), Mirror infers the unmeasured quantity from the physics model constrained by surrounding sensor readings — creating “virtual sensors” with quantified uncertainty
1:1
Sensor-to-model node pairing
Multi
Physics domain alignment
Virtual
Sensor inference where physical impossible
Gate
Data quality filtering before calibration
03
Model Calibration & Drift Detection
Bayesian parameter updating · Drift classification · Recalibration triggers · Model version governance
▼
The most valuable signal a digital twin produces is not its prediction — it is the moment when its prediction diverges from reality. Divergence means the physical system has changed in a way the model does not capture. Mirror continuously compares twin predictions against sensor readings and classifies divergence into three categories: random noise (within expected uncertainty bounds — no action), systematic bias (consistent offset requiring model recalibration), and progressive drift (growing divergence indicating physical degradation). Systematic bias triggers automatic Bayesian parameter updating — adjusting the physics model’s calibration parameters to minimize the residual. Progressive drift triggers a maintenance advisory and links to Trace for root cause analysis.
Bayesian parameter updating: When systematic bias is detected, Mirror identifies which physics model parameters (material properties, damping coefficients, boundary conditions) most likely explain the observed deviation and updates them using Bayesian inference — preserving the physics structure while adapting to reality
Drift classification: Three-class classifier distinguishes random noise (no action), systematic bias (recalibrate model), and progressive drift (physical degradation — maintenance advisory). Classification prevents false alarms from sensor noise while catching genuine degradation signals
Model version governance: Every calibration update creates a new model version in Nexus, governed by Sentinel’s audit trail. The twin’s calibration history is traceable: which parameters changed, when, based on what data, and by how much. Regulatory-grade model provenance
Bayes
Parameter updating from observed data
3
Drift classes (noise, bias, degradation)
Auto
Recalibration on systematic bias
Gov
Model version audit trail via Sentinel
04
Remaining Useful Life Prediction
Physics-based fatigue curves · Unit-specific load history · Damage accumulation · Maintenance window optimization
▼
Calendar-based maintenance wastes money. Condition-based maintenance wastes less. Physics-based predictive maintenance wastes almost none. Mirror predicts remaining useful life by running the actual operating load history of each specific unit through the physics-based fatigue curves from Nexus simulation. Not fleet averages — the actual loads this specific unit has experienced, as measured by its IoT sensors. A unit operated conservatively at 80% rated load has more remaining life than one operated at 110% rated load for the same number of hours. Mirror knows the difference because it has tracked the actual load profile through the twin.
Miner’s rule damage accumulation: Cycles at each stress amplitude are counted from the actual load history and accumulated against the S-N fatigue curve from Nexus simulation. Total accumulated damage (D = Σni/Ni) predicts failure when D approaches 1.0, with statistical adjustment for material scatter
Multi-mechanism degradation: Fatigue is one mechanism. Mirror tracks multiple degradation mechanisms simultaneously: creep (for high-temperature components), wear (for contact surfaces), corrosion (for exposed materials), and thermal cycling (for electronic assemblies). RUL is governed by the shortest remaining life across all mechanisms
Maintenance window optimization: RUL prediction is translated into an optimal maintenance window that balances remaining life against production scheduling constraints, spare parts availability, and maintenance crew capacity. Mirror recommends when to maintain — not just that maintenance is needed
Unit
Specific RUL (not fleet average)
Multi
Mechanism degradation tracking
Miner
Physics-based damage accumulation
Window
Maintenance timing optimization
05
Anomaly Detection & Operating Envelope
Multi-parameter correlation · Autoencoder-based detection · Envelope exceedance tracking · Alert classification
▼
Traditional anomaly detection monitors individual parameters against thresholds. Mirror detects anomalies in the relationship between parameters — because degradation often manifests as a changing correlation before any single parameter exceeds its limit. A bearing running hotter than predicted at the same speed and load reveals degradation that threshold-based monitoring would miss until the temperature alarm fires. Mirror uses autoencoder networks trained on normal operating data to detect anomalous multi-parameter patterns that deviate from the learned normal state.
Autoencoder anomaly detection: Neural network trained to reconstruct normal operating patterns. When the reconstruction error exceeds a calibrated threshold, the input represents an anomalous operating state — even if every individual parameter is within its normal range. The anomaly is in the combination, not the individual values
Operating envelope tracking: Records the actual operating envelope of each unit: maximum and minimum values of every monitored parameter, and the time spent in each region of the operating space. Units that operate near envelope boundaries accumulate damage faster than those operating near nominal conditions
Alert classification: Anomalies are classified by severity (advisory, caution, warning, alarm), by degradation mode (thermal, mechanical, electrical), and by urgency (monitor, schedule maintenance, stop operation). Classification prevents alert fatigue while ensuring genuine threats receive immediate attention
Multi
Parameter correlation detection
AE
Autoencoder anomaly classification
Envelope
Operating region tracking per unit
4
Alert severity levels with routing
06
Predictive Maintenance Scheduling
RUL-driven scheduling · Production calendar integration · Spare parts coordination · Maintenance crew optimization
▼
Knowing that a bearing has 2,200 hours of remaining life is only useful if the maintenance can be scheduled at the right time. Mirror translates physics-based RUL predictions into actionable maintenance schedules that account for production calendar constraints, spare parts lead times, maintenance crew availability, and the cost tradeoff between early replacement (wasted remaining life) and late replacement (increased failure risk). The scheduling engine integrates with Forge ERP’s maintenance management module to create work orders with linked parts lists, tool requirements, and procedure references.
Cost-optimal scheduling: Balances the cost of premature replacement (wasted remaining component life) against the cost of unplanned failure (production loss + emergency repair + potential collateral damage). The optimal maintenance window minimizes total expected cost
Spare parts pre-staging: When RUL prediction indicates maintenance within a planning horizon, Mirror triggers spare parts procurement through Forge ERP — ensuring parts are on-site before the maintenance window opens. No maintenance delays waiting for parts
Multi-asset coordination: When multiple assets on the same production line approach maintenance windows simultaneously, Mirror coordinates the schedules to minimize total production downtime — grouping maintenance activities during planned shutdowns where possible
Cost
Optimal maintenance window calculation
ERP
Forge work order auto-generation
Parts
Pre-staging via procurement pipeline
Fleet
Multi-asset schedule coordination
07
Closed-Loop Design Feedback
Fleet-wide deviation patterns · Simulation assumption correction · ECR auto-generation · Model evolution tracking
▼
The most powerful capability of a digital twin is not predicting the future of a single unit — it is revealing systematic gaps between design assumptions and field reality across the entire fleet. When Mirror detects that every unit in the fleet runs 10% hotter than the thermal model predicts under the same operating conditions, the deviation is not unit-specific degradation — it is a systematic gap in the simulation model’s boundary conditions or material properties. Mirror aggregates deviation patterns across the fleet, identifies systematic biases, and feeds them back to the design team via Cascade as corrective ECRs. The simulation model in Nexus is updated to reflect field reality — and the next product generation benefits from more accurate design-time predictions.
Fleet deviation aggregation: Individual unit deviations are noisy. Fleet-wide patterns are signal. Mirror aggregates the prediction-vs-reality delta across all twins of the same product configuration to identify statistically significant systematic biases that affect the entire fleet
Simulation assumption correction: When a systematic bias is identified, Mirror traces it to the specific simulation parameter (boundary condition, material property, load assumption) that most likely explains the fleet-wide deviation — and proposes the corrected value based on the aggregated field data
Corrective ECR via Cascade: Fleet-wide design feedback generates a draft ECR in Cascade with the aggregated evidence, proposed simulation correction, and impact analysis. The loop from field → twin → simulation → design is formally closed through the governed change management process
Fleet
Wide deviation pattern analysis
Sim
Parameter correction from field data
ECR
Auto-generated via Cascade pipeline
Loop
Field → Twin → Design (closed)
08
Fleet Twin Intelligence
Multi-unit orchestration · Cross-fleet pattern recognition · Operating regime clustering · Benchmarking analytics
▼
A single digital twin monitors one asset. A fleet of digital twins generates intelligence that no single twin can produce. Mirror’s Fleet Intelligence engine orchestrates hundreds or thousands of individual twins, comparing their degradation trajectories, identifying outlier units, clustering operating regimes, and benchmarking performance across sites, operators, and operating conditions. The fleet perspective answers questions no single twin can: “Which units are aging fastest?” “Which operating regime produces the longest component life?” “Which maintenance crew achieves the best post-maintenance performance?”
Degradation trajectory comparison: Plots the degradation trajectory (measured via twin divergence growth rate) of every unit in the fleet. Outlier units — those degrading significantly faster than the fleet average — are flagged for investigation via Trace. The comparison reveals whether rapid degradation is unit-specific or site-specific
Operating regime clustering: Groups units by their actual operating profile (load distribution, duty cycle, environmental exposure) rather than by product configuration. Units with similar configurations but different operating regimes may have vastly different remaining life — the fleet view makes this visible
Operator benchmarking: Different operators run the same equipment differently. Fleet Intelligence identifies which operating practices correlate with longer component life and lower maintenance costs — enabling best-practice dissemination across the installed base
1000+
Simultaneous twin orchestration
Outlier
Rapid degradation unit flagging
Cluster
Operating regime grouping
Bench
Operator practice benchmarking
DEPLOYMENT EVIDENCE

Three fleets. Prediction realized.

POWER GENERATION · GAS TURBINE FLEET
Digital twin fleet predicts bearing degradation 45 days before unplanned shutdown across 24 turbine units
24 gas turbines · 18 sensor channels per unit · Physics-hybrid twin architecture
A power generation operator running 24 gas turbines deployed Mirror twins with 18 sensor channels per unit (vibration, temperature, pressure, speed, exhaust gas analysis). Within the first 6 months, the vibration twin on Unit GT-7200-A detected a progressive drift — physical vibration exceeded model prediction by 35% while temperature deviation was only 3%. The divergence pattern matched bearing degradation, not rotor imbalance (which would produce correlated vibration and temperature drift). Mirror predicted 2,200 hours remaining useful life and recommended a 45-day maintenance window. The bearing was replaced during a scheduled outage, preventing an estimated $1.8M unplanned shutdown. Across the 24-unit fleet, Mirror predicted 7 maintenance events an average of 38 days before they would have resulted in unplanned downtime.
7
Maintenance events predicted proactively
38d
Average prediction lead time
$8.4M
Unplanned downtime cost avoided
AEROSPACE · ENGINE MRO · 340 ENGINES
Fleet twin intelligence identifies operating regime that extends hot-section life 22% across 340 engines
340 engines · 8 operating regime clusters · Fleet-wide degradation trajectory analysis
An engine MRO provider managing 340 engines across 12 airline customers deployed Mirror’s Fleet Twin Intelligence engine. Operating regime clustering identified 8 distinct usage profiles based on actual flight data (takeoff thrust setting, cruise altitude, derate percentage, ambient temperature distribution). Degradation trajectory comparison revealed that engines operating in the “high-derate, moderate-climb” cluster exhibited 22% slower hot-section degradation than the fleet average — despite flying equivalent cycle counts. The MRO provider shared the operating practice guidelines with all 12 customers. Within two service intervals, fleet-wide hot-section life increased by an average of 14%, reducing shop visit frequency and saving an estimated $12M annually across the managed fleet.
22%
Hot-section life extension identified
340
Engines in fleet twin network
$12M
Annual savings from optimized operations
INDUSTRIAL EQUIPMENT · COMPRESSOR FLEET
Closed-loop design feedback corrects 15% thermal underestimation in next-generation compressor design
180 compressor twins · Fleet-wide thermal deviation pattern · Simulation model corrected
A compressor manufacturer deployed 180 Mirror twins across their installed base. Fleet deviation aggregation revealed a systematic pattern: every unit ran 15% hotter than the thermal simulation predicted at rated operating conditions. The deviation was not unit-specific degradation — it was consistent across the entire fleet from commissioning. Mirror traced the systematic bias to the convection coefficient used in the thermal boundary condition of the Nexus simulation model, which had been derived from laboratory test data rather than field installation conditions (confined enclosures with restricted airflow). A corrective ECR was generated through Cascade with the field-calibrated convection coefficient. The next-generation compressor design used the corrected thermal model — eliminating a design margin that had been compensating for the model error and reducing material cost by 8% while maintaining the same thermal safety factor.
15%
Systematic thermal bias corrected
180
Twins in fleet deviation analysis
8%
Material cost reduction (next-gen design)

“The vibration was 4.2 millimeters per second. The twin predicted 3.1. That 35% divergence told us something was changing inside the bearing housing that no single sensor reading would have flagged — because 4.2 is still below our alarm threshold. The twin saw the degradation pattern forty-five days before the bearing would have failed. Forty-five days. That is the difference between a planned replacement during a scheduled outage and a one-point-eight-million-dollar unplanned shutdown.”

Director of Asset Reliability
POWER GENERATION · 24 GAS TURBINE FLEET

“We discovered that every compressor in our installed base runs fifteen percent hotter than we designed it to. Not degradation. Not manufacturing variation. A systematic gap between our simulation model and field reality — because we derived our thermal boundary conditions from lab tests, not from the confined enclosures our customers actually install these machines in. One hundred and eighty twins told us the same story. We corrected the simulation model. The next generation design saved eight percent on materials while maintaining the same safety margin. That is the closed loop.”

VP of Engineering & Simulation
INDUSTRIAL COMPRESSOR OEM · 180 INSTALLED UNITS

Stop simulating
products at design time.
Start mirroring them
for life.

Connect your first IoT sensor stream. Watch Mirror build a physics-data hybrid twin, calibrate against reality, and start predicting what your simulation never could.

Or contact the Mirror engineering team at mirror@brindwell.com