Physicsbased models do not require large amounts of data but are generally limited by their computational complexity or incomplete physics. In contrast, machine learning models appear promising for complex systems that are not fully understood or represented with simplified relationships, given adequate quality and quantity of data. We investigate applications of machine learning techniques for a wide variety of complex phenomena. Our application examples include additive manufacturing, multiphysics dynamics problems, damage detection in concrete structures, air transportation system safety, rotorcraft operations, power grid reliability, and cancer patient safety.
Machine learning (ML) models and strategies pursued:
Models
 Supervised learning
 Regression

Gaussian process (GP) models

Deep neural networks (DNN)

Bayesian neural networks (BNN)

Support vector machines

 Classification
 Knearest neighbor algorithm
 Regression
 Unsupervised learning
 Dimension reduction
 Principal component analysis (PCA)
 Clustering
 Kmean
 Dimension reduction
 Reinforcement learning
 Convolution neural networks (CNN)
 Long shortterm Memory (LSTM)
 Convolutional LSTM
 Graph neural network
Strategies
 Active/adaptive learning
 Ensemble/hybrid learning
 Transfer learning
 Physicsinformed machine learning (PIML)
We use machine learning for the following purposes: building surrogate models for expensive physics models, constructing datadriven machine learning models, learning multivariate probability distributions, reducing the dimension of the problem, correcting the physics/datadriven model, and transferring information (model errors and discrepancies) from tested configurations to untested configurations.
Machine learning surrogates for realtime risk assessment of power grids (risk learning)
 Developed machine learning (ML) surrogates for power grid decisionmaking (optimization) algorithms; used a novel hazardaware loss function for model training
 Directly predicted quantities of interest (load shed, reserve capacity, etc.), which depend on the decisions (generator commitment and dispatch) made by solving multiple constrained optimization problems
 Developed a justintime learning methodology, where ML surrogates for each day are trained every day and continuously updated during the dayahead market settlement and reliability assessment phase
 Used ML surrogates for realtime risk estimation given the latest (e.g., three hoursahead) forecast of load and wind/solar generation, and improved grid operator’s situational awareness
Machine Learning for Additive Manufacturing
 Gaussian process (GP) surrogate models for output quantities of interest (e.g., porosity, geometric accuracy, surface roughness, residual stress)
 GP surrogate models for model discrepancy, to correct the model prediction, trained using multiphysics simulations (fluid, thermal, mechanical).
 Computationally expensive highdimensional multiphysics models are used for the simulation of laser powder bed fusion (LPBF) additive manufacturing.
 Computational efficiency in the construction of surrogate models is achieved by using dimension reduction techniques such as active subspace, principal component analysis (PCA), and singular value decomposition (SVD).
 Deep neural network (DNN) to model the bond formation process and mesostructured of printed fused filament fabrication (FFF) parts.
 Physicsinformed machine learning (PIML) to improve the prediction accuracy and physics consistency of machine learning models.
Extrapolation of dynamics multiphysics models to untested conditions
Applying ML tools (e.g. artificial neural networks) to
 Create a hybrid physicsML model, where ML is used to correct model form errors in the physics model
 Create surrogate ML counterparts of physics models of dynamic systems à use for nonintrusive evaluation of model errors
 Use ML models à transfer information (model errors and discrepancies) from tested configurations/locations/load cases to untested prediction configurations/locations/load cases.
Ensemble machine learning for aviation incident risk prediction
 Aviation safety reporting system (ASRS) data
 Support vector machine (SVM) model to obtain risk category from event synopses
 Deep neural networks (DNN) to model the complex relations between event contextual features and event outcomes
 Fusion of two machine learning models to perform risk categorization
Multifidelity machine learning for enroute safety assessment
 FAA – system wide information management (SWIM) data for flight trajectories
 Deep neural network (DNN) for onestepahead flight trajectory prediction
 Deep Long ShortTerm Memory (LSTM) neural network for longerterm flight trajectory prediction
 Multifidelity prediction by blending the two deep learning models
Bayesian deep learning for aircraft hard landing safety assessment
 DASHlink data on entire flight trajectory à extract landing information
 LSTM neural network for early warning of hard landing
 Bayesian neural network to capture epistemic uncertainty in model prediction
Bayesian network for aviation accident risk prognosis
 Analyze the historical passenger airline accidents that happened from 1982 to 2006 as reported in the National Transportation Safety Board (NTSB) aviation database.
 Develop a Bayesian network representation of all the accidents by capturing the causal and dependent relationships among a wide variety of contributory factors and event sequences in terms of aircraft damage and personnel injury.
 Estimate the conditional probabilities in the Bayesian network by calibrating a monotonically increasing function, whose parameters are calibrated using the probability information on single events in the available data.
 Automate the generation of the Bayesian network in compliance with the XML format used in the commercial GeNIe modeler.
Airport ground safety assessment
 SWIM Terminal Data Distribution System (STDDS) for surface movement event data
 Spark for constructing sequences of graphs to represent the trajectories of multiple objects on the airport ground
 Graph neural network for predicting the probability distribution of future ground movements
 Conflicting risk assessment using the model prediction
Damageadaptive rotorcraft maneuvering
 Diagnosis, prognosis and decisionmaking for rotorcraft (inspection/maintenance/repair, mission planning, operational maneuvers).
 Training of machine learning models (Gaussian process (GP), DNN, LSTM, reinforcement learning) using rotorcraft dynamics and finite elementbased stress analysis models.
 Sensor data used for damage diagnosis and model calibration.
 Mission planning: Optimize future mission profile for safe mission completion, based on diagnosis of current vehicle state and prognosis of future damage growth.
 Damage adaptive maneuver: Adjust flight maneuvers to minimize vibratory loads and redistribute component stress, in order to minimize damage growth.
Concrete Damage Diagnosis with Vibroacoustic Testing
 Supervised machine learning using training data constructed from 2D FEA models for concrete structures with a single hidden crack.
 Two datadriven models constructed for damage diagnosis.
 Prediction model that estimates the sideband sum (SBSum) at a sensor location, from vibroacoustic modulation (VAM) testing. Bayesian damage diagnosis localizes the damage using observed VAM test data.
 Classification model that uses VAM test parameters and the measured SBSum values at a particular sensor to classify the locations as damaged or undamaged.
Concrete Damage Diagnosis with Harmonic Vibration Testing
 Damage detection with damage sensitive features identified by Singular Value Decomposition (SVD)
 Support Vector Machine classification model.
 Damage localization with crest factor and kmeans clustering.
Concrete damage diagnosis based on infrared thermography
 Augment experimental data with 2D and 3D finite element models.
 Damage localization and quantification with transfer learning from sophisticated deep neural networks, such as VGG, ResNet, etc.
 Uncertainty quantification of damage diagnosis using Monte Carlo dropout.
Cancer patient safety
 Data collection from cancer patients through
 Passive surveillance with wearable sensors (activity, heart rate, and geolocation data)
 Active surveillance with mobile appbased patient/caregiver reporting of nonroutine events and patientreported outcome measures
 Construction of predictive model of clinical deterioration in patient using machine learning methods such as random forest, artificial neural networks, etc. to predict the possible occurrence of unplanned treatment events
 Enabling timely response to clinical deterioration through early detection and prediction of clinical deterioration
Funding
Current People
 Sankaran Mahadevan, Professor
 Pranav Karve, Assistant Research Professor
 Xiaoge Zhang, FedEx
 Abhinav Subramanian, Postdoctoral Research Scholar
 Yanqing Bao, Research Engineer
 Berkcan Kapusuzoglu, Ph.D. Student
 Yulin Guo, Ph.D. Student
 Sarah Miele, Ph.D. Student
 William Sisson, Ph.D. Student
 Sanqiang Zhong, M.Eng. Student
 Paromita Nath, Postdoctoral Research Scholar
Publications
 Kapusuzoglu B., Sato M., Mahadevan S., Witherell P., “Process Optimization under Uncertainty for Improving the Bond Quality of Polymer Filaments in Fused Filament Fabrication”, Journal of Manufacturing Science and Engineering, 2020 Aug 18:146
 Kapusuzoglu, B., Mahadevan, S. PhysicsInformed and Hybrid Machine Learning in Additive Manufacturing: Application to Fused Filament Fabrication. JOM 72, 4695–4705 (2020).
 Zhang X., and Mahadevan S. "Bayesian neural networks for flight trajectory prediction and safety assessment." Decision Support Systems 131 (2020): 113246.
 Zhang X., and Mahadevan S., "Ensemble machine learning models for aviation incident risk prediction." Decision Support Systems 116 (2019): 4863.
 Subramanian, A., & Mahadevan S. “Bayesian estimation of discrepancy in dynamics model prediction.” Mechanical Systems and Signal Processing 123 (2019), 351368.
 Subramanian A., & Mahadevan S. “Model Error Propagation in Coupled Multiphysics Systems.” AIAA Journal 58(5) (2020), 22362245.
 Bao Y., and Mahadevan S., "Harmonic vibration testing for damage detection and localization in concrete." Structural Health Monitoring 18.56 (2019): 18201835.