Machine Learning
Don’t Blink. AI in Industry and Future Opportunities ARPA-e Workshop June 21, 2018 Michael Giering Technical Fellow: Machine Intelligence & Data Analytics United Technologies Research Center East Hartford, CT
DL - Deep Learning AI in Industry and Future Opportunities ML - Machine Learning AI – Artificial Intelligence Outline Machine Learning What is it Recent progress Current State of DL for Engineering Industry Challenges for Design & Manufacturing Design and Manufacturing Opportunities for ML Conclusions 3
Machine Learning : “Learning Programs From Data” AI 2012 Decision making via rule Speech Recognition based and expert Object Recognition systems 2014 I nformation Fusion 2015 Machine Learning Sequential Decision Probabilistic methods that Making improve with more data Attention Models 2016 Generative Adversarial Networks Deep Learning Text -> I mage Creates the best data 2017 representations to date for Unsupervised learning and querying. Generative Models Meta-Learning 2018 Self Attention Cross Domain Modeling 4
Current State of Deep Learning for Engineering Awaiting Newton , though Copernicus would do Ptolemy Copernicus Newton 1. Also consistent with the data. 1. This is what the data tells us. 1. Also consistent with the data. 2. The underlying principle. 2. Explainable. Comforting. 2. Best available predictions. 5
Industry Challenges for Design & Manufacturing Common Industry Practices Machine Learning Shortcomings Physics based engineering Difficulty incorporating explicit constraints Physics, manufacturing & engineering specs Reliance on extreme scale simulations Heuristic, incremental design methods Expert based decision making Risk and Bottlenecks Many expert based tasks: Produce highly variable results. Are repetitive, time consuming and unscalable. Pace Are difficult to codify. The rate of ML innovation is several times the rate of non-software product Organizational innovation. Often suboptimal: Data collection and management. Competitive advantage is fleeting. Analytics planning and pipeline standardization. Data collection and feedback loop for design. 6
Design and Manufacturing Opportunities for Machine Learning Learning from Prior Design & ML-enabled Material Design and Characterization Multi-fidelity Design Optimization Detect when massively parallel simulations can be modeled at lower fidelity and switch. Design of Complex Energy System Components Spec Consistent Automated High-dimensional Design The power of ML Unsupervised & Supervised Learning Constrained to: Advanced Manufacturing Physics, Engineering Constraints and Manufacturing ML-enabled large dimensional multi- Heat exchanger design recommendations, Representation Learning specs disciplinary design fabrication and validation 7
Bridging the Engineering – Machine Learning Gap Conclusions The greatest barrier to realizing the value of ML is the absence of bridges from existing deterministic, rule based practices to highly non-linear probabilistic methods. Trust by experts Performance confidence and explainability Cutting edge DL methods are Validation methods becoming more explainable. Specification practices Generative models are enabling better Commissioning practices and more generalizable models. Unsupervised learning has begun and is the key to exploration of design and manufacturing product spaces. 8
Thank You Michael Giering GierinMJ@ utrc.utc.com
Recommend
More recommend
