SBIR-STTR Award

Process industries have a pressing need to assess the state of materials at each stage in their supply chains – starting with raw materials at a vendor location and ending with the finished goods at customer sites. Such assessments entail the measurement of material properties like chemical composition and physical properties. Today’s assessments (done in expensive analytical laboratories) are largely based on cumbersome, intermittent manual sampling, which is expensive, time consuming, and error-prone. This project addresses DOE’s interest in turn-key solutions advancing the use of HPC in manufacturing supply chains. Our innovative idea is to use ML (machine learning) algorithms running on HPC servers to process spectra signals gathered from handheld infrared sensors for rapid and frequent material identification/estimation in supply chains. Our overall Phase I and Phase II objective is to bring to market a cutting-edge, HPC-based material identification application which produces accurate near real-time readings. The confluence of HPC, ML, and handheld sensors is at the center of our innovative approach and our project will be the first one to accomplish this. We will leverage the extensive expertise in ML, material sensing, and spectral data analysis from Rutgers University, which is our partner in this STTR project. In Phase I, we will demonstrate feasibility by developing ML algorithms which use data from handheld off-the-shelf IR sensors for material identification and prove the concept on a HPC test-bed. We will leverage our relationships with our industrial partners Constellation Brands (a large wine, spirits, and beer producer), Johnson & Johnson (a pharmaceutical co.), and Viavi (maker of IR sensors) to address real world material identification challenges (letters of support are attached). We will focus on estimating product quality at the blending location in a large winery site in our Phase I work. We will demonstrate end-to-end performance on our test-bed: data gathering, ML, and the use of the resultant models for quality estimation. Commercial Applications & Other

Benefits:
The envisioned commercial application will be used to support fast decision-making in supply chains to produce better quality & safer products. It is projected that companies will save millions of dollars by reducing supply chain costs and product recalls.

Material analysis is critical to enable the production of safe and quality goods and today companies use cumbersome, time-consuming, and error-prone methods to analyze materials using expensive analytical laboratories. Companies have a pressing need to measure material properties (like chemical composition and physical properties) throughout their supply chains. This project addresses DOE’s interest which is to bring the benefits of high performance computing (HPC) in manufacturing supply chains via turn-key solutions.Overall Objective and Approach: Our overall Phase I and II objective is to bring to market a cutting-edge material analysis application which produces accurate, near real-time readings. Our innovative idea and approach is to use machine learning (ML) algorithms leveraging HPC to process data gathered from handheld infrared sensors for rapid material analysis. Rapid material analysis enables companies to produce quality products safely. The confluence of HPC, ML, and handheld sensors is at the center of our innovative approach and our project will be the first one to accomplish this.Phase I R&D resulted in the development of (1) ML-based calibration models for portable sensors and (2) the technology to parallelize the training of the ML models. We demonstrated the feasibility of using HPC and portable sensors to achieve 80-times calibration speedup, 10,000-times faster analysis cycle time, and 70% cost reduction (vs. legacy lab-based sensors) without a significant loss of accuracy. We also demonstrated a latency of less than 1 second per sample analysis which enables near real-time analysis and control. Our results were reviewed by (1) prospective customers who gave us validation and encouraging feedback (see support letters) and (2) industry peers who approved our results for presentation at Pittcon-2018.Phase II work will expand our Phase I models to address an extended list of requirements (like dissolution rates of tablets, counterfeit drug detection, etc.). We will also focus on overcoming challenges including adaptation of our product for in-process use, scalability, and data security. Our target is to be ready by the end of Phase II with a prototype turnkey application running on a hosted-HPC infrastructure. Commercial Applications: Our product will be used to produce better quality and safer products. Companies will potentially save millions of dollars from reductions in supply chain costs and product recalls. Prospective customers include the FDA-regulated wine & spirits and pharmaceutical companies. Per FDA, companies recall thousands of food and drug products annually. Timely material analysis is key to reducing recalls. This presents us with a huge market opportunity.