Integrated Modeling, Inference, and Computing

Integrated Modeling, Inference and Computing will focus on the advancement of the integration of core areas of engineered modeling approaches, machine learning and computation to address barriers in smart modeling with applications in bioengineering for health & disease, environmental health monitoring & climate change, and engineering & design of advanced material systems. It will identify testbeds that define broad application areas that demand new developments in our three fundamental core areas to address barriers in smart modeling.

Northeastern University has significant strengths and expertise in computational modeling, machine learning, and computing.  While each of these areas represents a rich technology base, the future lies in their integration to allow the development of smart modeling of complex systems and quantitative model-based learning that can leverage compute-enabled discovery to truly take advantage of the explosion of available data.  The concept of smart systems suggests that models can “learn”, and leverage this knowledge to create systems that can tune their parameters, both within and across scales; add new model components; as well as profile and optimize their behavior to provide new levels of specificity and accuracy. Model-based learning expands data exploration methods beyond purely statistical approaches to use tailored computing methods to incorporate domain knowledge into data analysis.  We envision bold breakthroughs in the areas of material design, environmental monitoring, and human health.

The initiative will focus on fundamental advances to the integration of the core areas of modeling approaches, machine learning, and computation.  To motivate these advances, we identify testbeds that define broad application areas that demand new developments in our three fundamental core areas to address barriers in smart modeling.  Initial system-level applications that will motivate our research include: designing new nano-materials that comprise interactions at multiple scales and dimensions, understanding malignant cancerous cell behavior, integrating psychological and mechanistic models of happiness and learning with real-time measurements, and characterizing environment factors that result in drastic climate change.  These are representative of a compelling set of applications that remain major challenges to science and engineering.

The modeling challenges in these applications inherently involve huge volumes of data.  Smart models require new machine learning techniques to adapt and expand mechanistic models to fit complex phenomena. Model-based inference requires new methods of modeling to frame and guide data analysis. Both require advances in computing and algorithm rethinking to achieve reasonable time and memory burdens and to move from the laboratory into the real world.  Efficient computation is a critical enabler in all these approaches, and innovations are sorely needed to address the scale of the processing required in modeling spatial, temporal and networked dynamic systems, and seamless integrating processing from sensors, distributed embedded processors, heterogeneous complex networks and heterogeneous processing facilities. Our model of computation includes software and hardware at all these levels and involves data structures, algorithms, architectures and application-specific devices. We view each of these components as an integrated, complex, processing architecture, which can support the integrated modeling and learning we propose.

Specific applications will be used to advance modeling, machine learning, and computation.  The goal is unifying approaches that span these applications.  Some applications which researchers at Northeastern have established expertise, but continue to face significant computational barriers include:

Engineering and design of advanced material systems:  COE has expertise in scale bridging efforts, from atoms to continua, aimed at understanding structure-property relations in materials and their rationale design.  There is an urgent need for novel and highly efficient algorithms that allow seamless handshaking of the scales to make contact with experimentation; an integrated approach that involves heterogeneous computing and machine learning can pave the way for transformative approaches that have an immediate impact for quantifying the materials genome.  Systems of interest include structural materials such as metals and ceramics, electronic and photonic materials, nanomaterial behavior and manufacturing, material mechanics, and catalysis.

Bioengineering for health and disease: Human health involves complex interacting systems at spatial scales from molecular to societal and time scales from nanoseconds to decades. Full mechanistic models are infeasible, while generic statistical learning ignores known constraints (genetic, biophysical, environmental, social).  Advances from genetics and imaging to smartphones provide an avalanche of health-related data. We will integrate modeling, learning, and computing to extract information from these data for increased understanding of health and disease processes, eventually leading to fundamental health improvement. As examples, blood-borne sensors monitoring low-level physiological processes communicating with sophisticated smart models and learning algorithms could detect and treat early-stage metastasis or infection, and real-time monitoring of autonomic or non-conscious mental phenomena could allow entirely new methods of behavioral and psychological therapy.

Environmental health monitoring and climate change: From fundamental advances in the science of climate change and associated weather extremes to preventing hazards from turning into catastrophic disasters through improved resilience of critical infrastructures and sensor-based early warning systems, computational modeling and analysis have had increasingly large roles in societal priorities.., Machine learning is barely developing methods for handling extreme values or complex dependence structures, while high-performance computation is struggling to keep up with the volumes and velocity of data from models and sensor systems. COE has recognized capabilities in climate extremes,  hydrology, environmental health, life cycle analysis, sensor-based monitoring in geosciences, and critical infrastructures assessment.