Friday, January 22, 2021

U.S. Army will use AI to develop chemical-biological defense systems

The U.S. Army has announced plans to use Artificial Intelligence and Machine Learning (AI/ML) potential in the development of chemical and biological defense solutions.

According to a press release issued Monday by the U.S. Army Combat Capabilities Development Command Chemical Biological Center, AI/ML is a next-generation computer capability that holds the potential for changing everything from how people live and work to how wars are fought and won.

Also noted that the Combat Capabilities Development Command Chemical Biological Center (DEVCOM CBC) is keeping pace with this computing revolution through its Grand Challenge Program, beginning with three pilot projects and a workshop to recruit more.


“We are starting with the three pilot projects to demonstrate the value of AI/ML across the Center,” said Patrick Riley, a Center research chemist who is heading up the effort. “Our goal is to start up lots of small projects on the ‘fail fast’ principle so we can see what the best uses of AI/ML are here at the Center.”

Patrick Riley, who is leading the Chemical Biological Center’s artificial intelligence and machine learning initiative known as STEWARD, coding a software program which uses AI/ML to advance his research. Photo by Jack Bunja

The program will focus on four key areas: education, small projects, infrastructure tools, and communication. The objective is to build a foundational understanding of AI/ML within the workforce, and to create opportunities for workforce participation in the form of hands-on projects.

The effort has been dubbed STEWARD, which stands for Software Tools and Educating the Workforce to use Artificial Intelligence for Research and Development. It is the product of the latest Grand Challenge, a contest held every three years at the Center to invite scientists and engineers to propose a bold new initiative for funding and development. Grand Challenge winners keep the Center at the forefront of innovative science. One past winner combines materials and biological sciences to create brand new materials that are biologically active. Another past winner advanced the Center’s ability to detect chemical and biological agents from just a few inches away to up to 24 meters inside a secure laboratory.

This latest leap into the future meshes with a larger Army Futures Command (AFC) effort to create an AI-ready workforce by making AI education available to all AFC employees. This, in turn, falls in line with the Defense Department’s total approach to AI, which is to make it an ingrained warfighter skill. Military planners expect that ultimately AI/ML will be the key to getting inside an adversary’s decision loop in battle. If AI/ML can make U.S. forces quicker at the observe–orient–decide–act process, known as the OODA loop, they can keep an adversary off balance and win through greater agility rather than through greater force. AI/ML also holds great potential for advancing the nation’s Chemical Biological Defense Program.

The STEWARD AI Program is run by a committee of five Center scientists, each with a different expertise but all keen to serve the Center’s mission to accelerate discovery and deliver new capabilities to the warfighter. It kicked off with a lecture series in which members of the workforce and several outside guest lecturers explained AI/ML generally and how they are using it to enhance their own research. The committee also created a STEWARD AI Program website on the Center’s intranet. It includes curated videos, AI experiments, and links to free tools to try AI/ML out.

Take the task of using computer vision to identify whether an image is or is not an apple. An AI/ML algorithm requires that the computer assign the image a 0 for ‘no’ or a 1 for ‘yes’. For a neural network, a dataset of images makes a number of ‘yes’, ‘no’ predictions. The software program assesses its prediction rate, and then updates the weights to try to lower its prediction error so that it predicts more accurately the next time. A researcher sets the number of times this error updating occurs, hopefully close to when the accuracy is the highest. Also, the researcher can add more characteristics by taking photos of different kinds of apples and from many different angles to depict different color variations, skin textures, and irregularities. This adds more information to the dataset, which will help a final model perform more accurately.

A popular open source form of AI/ML that researchers at the Center often use is called Random Forest. It got that name because it uses multiple decision trees to make individual decisions and then, in effect, holds a vote among the decision trees in order to make a prediction. The more independent each tree is in its selection of characteristics and the weights placed on them, the better – the wisdom of crowds. Unlike Neural Networks, Random Forest can tell us which characteristics were most important in making its decision, opening the black box of AI/ML.

Random Forest is an open source algorithm that functions as an ensemble method, averaging the results of several decision trees to make a prediction.

If we want to complicate matters in the earlier example of identifying an apple, we can show the computer camera eye a tomato. Both are round, red varieties of fruit, but while differentiating between them is easy for us, it can be very hard for AI/ML. The answer – keep adding training data, let the algorithm add more trees, keep making predictions and learning from them through the software’s self-iteration function. In time, the algorithm will be able to very accurately distinguish between the two. In fact, there now exists commercially available algorithms that not only will not confuse an apple with a tomato, but can be used by farmers to autonomously select apples at the just the right stage of ripeness to pick and send to market. Another algorithm does the same for tomatoes.

Just as a journey of a thousand miles starts with one step, so the transition to an AI/ML-savvy workforce starts with a few pilot projects. One of the Center’s first three projects uses AI/ML to preserve and use the Center’s legacy data, the notes in decades of scientists’ notebooks, by scanning them and forming a repository. Another seeks to use AI to recognize crucial objects on the battlefield by developing an algorithm that operates with a commercially available augmented reality headset. The third one aims at using an algorithm to quickly examine a vast library of chemical compounds to find the compounds best suited for stimulating metals to produce electronic charges that can be used to break up chemical agent molecules.

The Center has old researchers’ notebooks containing handwritten data going back decades. Some of this data is particularly valuable for current research. A team consisting of a Center mathematician, Thomas Ingersoll, Ph.D., and a computer scientist, Pronoti Kundu, Ph.D., selected a trove of notebooks full of data recording the results of a unique, never-to-be-repeated experiment. In the 1990s, the Army went through a post-Cold War lull that made a group of over a thousand Soldiers available to participate in an exercise in which one group wore full battlefield chemical biological protective gear and another did not. The two groups were then timed as they performed typical battlefield infantry, artillery and armor functions such as frontal assaults, precision fire, and logistics.

The data derived from these experiments were meticulously recorded in notebooks and serve as a goldmine of information to this day. But all of it was typewritten and smudged with fingerprints, coffee stains, and whiteout, making it unsuited for simple scanning. Ingersoll and Kundu decided they would find the best way to make the data available to current researchers. Kundu is pursuing an AI/ML-enhanced scanning approach using computer vision. She tried different AI/ML software programs to find the one best able to overcome these obstacles. That meant it had to be able to recognize tables in the first place, extract from them, and not record missing values as gibberish.

Meanwhile, Ingersoll is using the original digitalized data from the experiment, recorded in a now defunct version of Microsoft Access, to recreate the tables in a contemporary software system. He had to learn how to write software queries in the antiquated Access code and run a statistical tool known as analysis of variance (ANOVA) for post-retrieval processing. The process proved painstakingly slow but highly accurate. Kundu’s process was also very slow at first, but once she identified the best AI/ML software for the job, and set up a data entry and retrieval system, it became much faster, and almost as accurate as Ingersoll’s method.

When they finish the project they expect to be able to provide the Center with tested and verified recommendations on how best to preserve the Center’s rich repository of legacy data.

Center researchers have been using augmented reality as a teaching tool for the past several years. A commercially available augmented reality headset can use software to overlay graphics onto the real world through the glasses. A viewer can then look at an actual piece of equipment and have a software-created or an augmented reality (AR) overlay appear in the form of instructions on how to operate it. The team of Jacob Shaffer, a computer scientist, Don Lail, a multimedia specialist, Charles Davidson, Ph.D., a senior research scientist, and Gary Kipler, Ph.D., a physicist, want to add object recognition using AI/ML to teach the operation of chemical biological detection equipment and recognize chemical biological threats in the field.

This entails using a machine learning algorithm and ‘training’ it to recognize objects in its environment. For example, suppose the headset can identify the Army’s current chemical agent detection workhorse, the Joint Chemical Agent Detection Device, or JCAD. In that case, it can then draw upon a pre-trained library to show a warfighter in the field exactly how to use it. Ultimately they would like the headset to tell the warfighter wearing it that it sees nearby chemical or biological threat. If successful, this would be a significant contribution to warfighter safety and capability in the field.

Formally known as plasmonic catalysis, the idea is to find the ideal material to use to form an electron cloud by subjecting it to light. These electrons then enter a semiconductor in the vicinity, and if that semiconductor is in contact with a chemical agent, the excited electrons break up the agent’s molecular structure, rendering it harmless. This is where the team of Matthew Browe, a Center chemical engineer, and John Landers, Ph.D., a Center research scientist, see the opportunity to use AI/ML. The right algorithm can potentially identify new plasmonic catalyst materials that can provide practical, in-the-field decontamination. A successful pairing of a semiconductor with the newly-discovered plasmonic catalyst could be used in coatings, or even as a layer in a protective suit.

To get to that point, Browe and Landers need to find the best possible inorganic compound for generating excited electrons when exposed to light. As it happens, the University of California at Berkeley has developed a database, formally known as the Materials Project, with more than 130,000 possible compounds. The database was created through crowdsourcing within the scientific community. Browe and Landers are developing an algorithm that can search through this vast database and select materials known as high-performing plasmonic catalysts. They will then populate their own smaller dataset containing the electronic structure data of these materials to identify key electronic structure features common to all high-performing plasmonic catalyst materials.

They plan to use the findings of the model produced from the best plasmonic catalyst candidates within the Materials Project database to potentially identify new materials with similar electronic features which may serve as promising new, undiscovered plasmonic catalysts. Center researchers can use this smaller list for targeting synthesis and performance measurements in the laboratory. This will involve some trial and error, but unlike Thomas Edison trying out thousands of compounds before discovering tungsten for lightbulbs, results from a machine learning model such as the one being developed by Browe and Landers will yield a subset of candidates in the range of only dozens of trials. This is especially important now because pandemic restrictions have reduced available laboratory time, so optimal down-selection of materials prior to benchtop efforts is more important now than ever.

On Nov. 3, the STEWARD AI Program Committee held a workshop open to any Center researcher interested in seeing if AI/ML could enhance his or her data analysis in the course of their research. It was held over Microsoft Teams. In a process much like speed dating, more than a dozen researchers each had an opportunity to enter a Teams chat room with one or more members of the committee to see if that researcher’s data was a match for what AI/ML can do. Most of the participants went into it knowing relatively little about AI/ML but eager to find new and better tools to advance their research.

The research efforts presented to the committee members as candidates for funding as a new AI/ML pilot project varied widely. One researcher wanted to see if AI/ML could help her optimize how she constructs metal-organic framework catalysts. These are custom-made molecules built much like an erector set using organic struts and metallic nodes, creating void spaces for chemical warfare agent molecules to enter and be trapped and neutralized. Academic and defense laboratories throughout the world have made increasingly sophisticated MOFs over the years and expanded the functions they can perform. The range of possibilities for constructing MOFs are vast. They are based on the researcher’s selection of organic struts and metallic nodes with specific specialized functions in mind. AI/ML holds the potential for examining the full range of existing MOFs and their functions as a data set and making predictions on how to better select nodes and struts for enhanced functionality.

Another technology presented at the workshop involves a bio-identifier that uses fluorescence to indicate the presence of a threat agent. The researcher was interested in using AI/ML to predict the likelihood of a positive detection, or hit, for various threat agents then using AI/ML to fine tune the thresholds for detecting those agents in an effort shave seconds off of the detection time and reduce the system’s reliance on a human operator.

Yet another area of research brought to the workshop involves using AI/ML to better process the mountain of data that liquid chromatography mass spectrometers produce. Known as LC/MSs, they are a widely used laboratory analytical instrument for detecting proteins, lipids, and small molecule metabolites within biological samples at low concentrations. This allows researchers to model host response or mechanisms of action to their presence in the body. At the Center, this is crucial for gaining a better understanding of human exposure to chemical and biological agents. The results come out in the form of graphs comparing each substance’s mass to its charge and signal intensity in a form called spectrum data. They inevitably include false positives which must be identified and discounted. AI/ML can potentially automate and fine tune this function reducing false positives and adding speed.

Because of AI/ML, the world is changing. Ultimately every task involving large amounts of data used to get a result will employ AI/ML to reduce or altogether remove the humans from the chore of making sense of that data. It may be the biggest computer innovation since the microchip. It will transform everything from self-driving cars to vaccine development to dangerous jobs such as bomb disposal. It is already making dramatic changes in banking security, climate change modeling, shipping logistics and many other large-scale human endeavors.

Virtually every aspect of defense technology will be fundamentally altered, too. “The Chemical Biological Center is intent on employing AI/ML to not only keep up with this pace of change, but to take a leadership role in using it to better defend warfighters and the nation from the many current and emerging chemical biological agent threats that a changing 21st Century world will present,” said the Center’s director, Eric Moore, Ph.D.

Over time, no area of the Center’s research and operations will be untouched by AI/ML. The way the Center’s leadership sees it, the future is here, and the STEWARD AI Program is merely here to jump-start its use.

If you wish to report grammatical or factual errors within our news articles, you can let us know by using the online feedback form.

If you would like to show your support for what we are doing, here's where to do it:

Executive Editor

About this Author

Colton Jones
Colton Jones
Colton Jones is technology editor for Defenсe Blog. He has written about emerging technology in military magazines and elsewhere.