The MIT Trust Data Consortium aims to provide people, organizations, and computers the ability to manage access to their data more securely, efficiently, and equitably, while protecting personal data from incursion and corruption. As we have moved from the analog world to the digital world, our data, security, and governance systems have not kept pace. This has created numerous issues ranging from data insecurity (such as the large-scale government and private sector data losses of recent years) to a widening digital divide between rich and poor, including the global disenfranchisement of over 1.5 billion people who lack legal identity.
Electrochemical energy storage is emerging as a critical technology to enable sustainable electricity generation by alleviating intermittency from renewable sources, reducing transmission congestion, enhancing grid resiliency, and decoupling generation from demand. While several different rechargeable batteries have been proposed for and demonstrated in these applications, further cost reductions are needed for ubiquitous adoption. As such, recent research has focused on the discovery and development of new chemistries. Though exciting, most of these emerging concepts only consider new materials in isolation rather than as part of a battery system. Understanding the critical relationships between materials properties and overall battery price is key to enabling systematic improvements. In this presentation, I will discuss an approach to mapping feasible design spaces for incipient energy storage systems through techno-economic modeling and to using this knowledge to identify critical pathways at an early stage in the research and development process. While redox flow batteries will be used as an exemplar technology, the methods to be described here are applicable to a wide range of electrochemical systems and envisioned applications.
The rapid, stable cycling of rechargeable batteries requires well-controlled phase transformations of the redox active materials in each electrode, between the charged and discharged states. In Li-ion batteries, common intercalation materials, such as graphite and iron phosphate, undergo phase separation (into Li-rich and Li-poor phases), which limits the power density and causes degradation. A general mathematical theory, supported by recent x-ray imaging experiments, will be presented that shows how phase separation can be controlled by electro-autocatalytic reactions. For Li-metal batteries, theoretical and experimental results will be presented for the stability of lithium electrodeposition, controlled by electrokinetic phenomena in charged porous separators.
Our laboratory focuses on the science and applications of nanocrystals, especially semiconductor nanocrystal (aka quantum dots). Our research ranges from the very fundamental to applications in electro-optics and biology. There is an ongoing effort to address the challenges of making new compositions and morphologies of nanocrystals and nanocrystal heterostructures, and new ligands so that the nanocrystals can be incorporated into hybrid organic/inorganic devices, or biological systems. We are collaborating with a number of biology and medical groups to design nanocrystal probes that meet specific challenges.
Accelerated penetration of distributed energy resources (DER) for power generation and demand response (DR), the notion of just-in-time flexible consumption, are enabling the transformation from the current power grid structure to a modernized, cyber-enabled grid. In order to carry out an efficient design of Transactive Systems, a tightly integrated design of wholesale and retail markets and pricing policies is needed that incentivizes end-users' participation, accommodates physical constraints, and enables global objectives through local and distributed decision-making.
This talk will outline how this integrated design of wholesale and retail markets can be carried out. The two streams of research investigations from our lab will be featured: one is a dynamic wholesale market mechanism with the ability to make decisions at multiple time-scales, and the other is a hierarchical architecture capable of achieving volt-var control in the presence of large penetration of DERs and DRs. We will discuss how the results from these can be combined to result in an overall hierarchical Transactive architecture for smart distribution grids.
While Google's mission is to organize the world's information, this information needs to made understandable and usable. We hear a lot about data, and worse, “big data,” but far too little about its meaning and how to make it approachable for the people who need to use it. The solution is to treat data as a design problem, where it can be addressed by starting with end users and working back to the data in all its messy complexity. We can only make progress if we first consider audience and context, forcing us to reformulate the questions at hand and to reconsider the technical decisions and approaches made behind the scenes.
Our clothes help define us, yet the fabrics we wear have remained functionally unchanged for thousands of years. Recent breakthroughs in fiber materials and manufacturing processes allow us to design and wear fabrics that see, hear, communicate, change color, and monitor health — heralding the dawn of a “fabric revolution.” Our mission at Advanced Functional Fabrics of America (AFFOA) is to lead the convergence of advanced technology into fibers (“Moore’s Law for fibers”) resulting in fabric products that deliver value-added services to the user (“Fabrics as a service”).
This talk introduces a new generation of machine learning methods that provide state of the art performance and are very interpretable, introducing optimal classification (OCT) and regression (ORT) trees for prediction and prescription with and without hyperplanes. This talk shows that (a) Trees are very interpretable, (b) They can be calculated in large scale in practical times, and (c) In a large collection of real world data sets, they give comparable or better performance than random forests or boosted trees. Their prescriptive counterparts have a significant edge on interpretability and comparable or better performance than causal forests. Finally, we show that optimal trees with hyperplanes have at least as much modeling power as (feedforward, convolutional, and recurrent) neural networks and comparable performance in a variety of real world data sets. These results suggest that optimal trees are interpretable, practical to compute in large scale, and provide state of the art performance compared to black box methods.
Machine learning is disrupting multiple fields of human endeavor: healthcare, transportation, finance, communications, etc. Materials design is no exception in this disruption. Data-driven approaches can access the information embedded in years of experiments, perform rapid optimization of high-dimensional experimental conditions and design parameters, or design new molecules automatically. The Gomez-Bombarelli group at MIT combines cutting-edge machine learning models on experimental data with automation in physics-based atomistic simulations (molecular dynamics, electronic structure) to rapidly design and optimize new materials in multiple areas, such as: inverse chemical design of small molecules (drug-like molecules that optimally bind biological sites, organic-light emitting diode emitters, and organic battery electrolytes); virtual discovery of soft materials (lithium-conducting polymers and OLED transport materials); and chemical reactivity in the condensed phase (zeolite design for catalysis and chemical and thermal stability of organic electronics). There is great interest in using machine learning as the connector between multiple time and length scales: from electronic structure, to atomistic molecular dynamics, to coarse-grained models.
Despite continuous advances over many decades, automatic speech recognition remains fundamentally a supervised learning scenario that requires large quantities of annotated training data to achieve good performance. This requirement is arguably the major reason that less than 2% of the worlds' languages have achieved some form of ASR capability. Such a learning scenario also stands in stark contrast to the way that humans learn language, which inspires us to consider approaches that involve more learning and less supervision.
In our recent research towards unsupervised learning of spoken language, we are investigating the role that visual contextual information can play in learning word-like units from unannotated speech. This talk will outline our ongoing research in CSAIL to develop deep learning models that are able to associate images with unconstrained spoken descriptions, and present analyses that indicate that the models are learning correspondences between associated objects in images and their spoken instantiation.