Speeding Up Soft Tissue Simulation with Machine Learning Based Metamodels
Abstract
Research problemThe Office of Naval Researchs Naval Force Health Protection has an interest in developing Digital Twins and enablin,g Digital Engineering. Ideally, it would be possible to synthesize data from medical scans, the medical record, and sensors to creat,e a subject-specific in silico model of every service person. Then, it would be possible to conduct subject-specific simulation for,applications such as tracking the onset of injury (e.g., TBI) or designing optimal protective gear (e.g., advanced helmet technology,). However, at present, subject-specific simulations will be prohibitively computationally expensive for practical use. The goal of,this proposal is to use machine learning techniques (i.e., metamodeling creating models of models) to overcome this anticipated bo,ttleneck. Here, we will focus on the initial problem of rapid simulation of blocks of soft brain tissue. This is a starting point to,wards future work in whole organ simulation.ObjectivesThe goal of this workis to develop methods for rapid simulation of brain tissu,e. There are two main components to the proposed work: 1)Create a benchmark dataset of mechanical simulations of brain tissue; and 2,)Develop machine learning based metamodeling techniques that produce high accuracy approximations of simulation results in fractions, of a second. Through these overarching objectives, we will be able to meaningfully step towards real time prediction of brain tissu,e deformation.Technical approachAs a starting point, we will use Finite Element Simulation conducted with the open source FEniCS lib,rary to run our brain tissue simulations. We will design a high fidelity simulation, and a corresponding correlated low fidelity sim,ulation. The low fidelity simulations will be used to augment our training dataset, and ultimately to showcase what may be possible,through leveraging low fidelity simulations in the future. In addition to releasing our dataset, we will also release all code under, the appropriate license via GitHub. To create our machine learning based metamodel, our starting point will be our recently propose,d MultiRes-Wnet architecture that consists of two connected UNets with Muti-Resolution Convolutional Blocks and ResPaths. We will re,lease our machine learning software under the appropriate license via GitHub. Once the metamodel is trained, we will also conduct pi,lot studies that take advantage of our ability to freely explore the simulation parameter space. For example, we will be able to ide,ntify best-case and worst-case loading conditions in an individual tissue geometry, and identify regions in the tissue where damage,is most likely.Anticipated outcomeThe first anticipated outcome of this work is a paper and corresponding open source software with,tutorials that describe how to design effective metamodels for spatially heterogeneous brain tissue. The second anticipated outcome,of this work is a roadmap for constructing and evaluating metamodels of brain tissue. We anticipate that the pipeline that we propos,e here will be generally applicable to other simulation types, in particular other simulations relevant to developing human Digital,Twins.Impact on DoD capabilitiesIdeally, this work should transition directly to the Office of Naval Researchs Naval Force Health P,rotection Digital Twin initiatives. Successful completion of this work will be a building block in that direction. Approved for Publ,ic Release
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Feb 08, 2022
- Source ID
- N000142212066
Entities
People
- Emma Lejeune
Organizations
- Boston University
- Office of Naval Research
- United States Navy