To synthesize high-performance materials for organic photovoltaics (OPVs) that convert solar radiation into direct current, materials scientists must meaningfully establish the relationship between chemical structures and their photovoltaic properties. In a new study on Science Advances, Wenbo Sun and a team including researchers from the School of Energy and Power Engineering, School of Automation, Computer Science, Electrical Engineering and Green and Intelligent Technology, established a new database of more than 1,700 donor materials using existing literature reports. They used supervised learning with machine learning models to build structure-property relationships and fast screen OPV materials using a variety of inputs for different ML algorithms.

Using molecular fingerprints (encoding a structure of a molecule in binary bits) beyond a length of 1000 bits Sun et al. obtained high ML prediction accuracy. They verified the reliability of the approach by screening 10 newly designed donor materials for consistency between model predictions and experimental outcomes. The ML results presented a powerful tool to prescreen new OPV materials and accelerate the development of OPVs in materials engineering.

Organic photovoltaic (OPV) cells can facilitate direct and cost-effective transformation of solar energy into electricity with rapid recent growth to exceed power conversion efficiency (PCE) rates. Mainstream OPV research has focused on building a relationship between new OPV molecular structures and their photovoltaic properties. The traditional process typically involves the design and synthesis of photovoltaic materials for the assembly/optimization of photovoltaic cells. Such approaches result in time consuming research cycles that require delicate control of chemical synthesis and device fabrication, experimental steps and purification. The existing OPV development process is slow and inefficient with less than 2000 OPV donor molecules synthesized and tested so far. However, the data gathered from decades of research work are priceless, with potential values remaining to be fully explored to generate high-performance OPV materials.

To extract useful information from the data, Sun et al. required a sophisticated program to scan through a large dataset and extract relationships from among the features. Since machine learning (ML) provides computational tools to learn and recognize patterns and relationships using a training dataset, the team used a data-driven approach to enable ML and predict diverse material properties. The ML algorithm did not have to understand the chemistry or physics behind the materials properties to accomplish the tasks. Similar methods have recently predicted the activity/properties of materials successfully during materials discovery, drug development and materials design. Prior to ML applications, scientists had generated cheminformatics to establish a useful toolbox.

Materials scientists have only recently explored the applications of ML in the OPV field. In the present work, Sun et al. established a database containing 1719 experimentally tested donor OPV materials gathered from literature. They studied the importance of programming language expression of the molecules first to understand ML performance. They then tested several different types of expressions including images, ASCII strings, two types of descriptors and seven types of molecular fingerprints. They observed the model predictions to be in good agreement with the experimental results. The scientists expect the new approach to greatly accelerate the development of new and highly efficient organic semiconducting materials for OPV research applications.

The research team first transformed the raw data into a machine readable representation. A variety of expressions exist for the same molecule comprising vastly different chemical information presented at different abstract levels. Using a set of ML models, Sun et al. explored diverse expressions of a molecule by comparing their predicted accuracy for power conversion efficiency (PCE) to obtain a deep-learning model accuracy of 69.41 percent. The relatively unsatisfactory performance was due to the small size of the database. For instance, previously when the same group used a larger number of molecules of up to 50,000, the accuracy of the deep-learning model exceeded 90 percent. To fully train a deep-learning model, researchers must implement a larger database containing millions of samples.

Sun et al. only had hundreds of molecules in each category at present, making it difficult for the model to extract enough information for higher accuracy. While it is possible to fine-tune a pre-trained model to reduce the amount of data required, thousands of samples are still necessary to accomplish a sufficient number of features. This led to the option of increasing the size of the database when using images to express molecules.

The scientists used five types of supervised ML algorithms in the study, including (1) back propagation (BP) neural network (BPNN), (2) deep neural network (DNN), (3) deep learning, (4) support vector machine (SVM) and (5) random forest (RF). These were advanced algorithms, where BPNN, DNN and deep learning were based on the artificial neutral network (ANN). The SMILES code (simplified molecular-input line entry system) provided another original expression of a molecule, which Sun et al. used as inputs for four models. Based on the results, the highest accuracy approximated 67.84 percent for the RF model. As before, unlike with deep learning, the four classical methods could not extract hidden features. As a whole, SMILES performed worse than images as descriptors of molecules to predict the PCE (power conversion efficiency) class in the data.

The researchers then used molecular descriptors that can describe the properties of a molecule using an array of numbers instead of the direct expression of a chemical structure. The research team used two types of descriptors PaDEL and RDKIt in the study. After extensive analyses across all ML models, a large data size implied more descriptors irrelevant to PCE affecting the ANN performance. Comparatively, a small data size implied inefficient chemical information to effectively train ML models, when using molecular descriptors as input in ML approaches, the key relied on finding appropriate descriptors that directly related to the target object.

The team next used molecular fingerprints; typically designed to represent molecules as mathematical objects and originally created to identify isomers. During large-scale database screening, the concept is represented as an array of bits containing “1” s and “0” s to describe the presence or absence of specific substructures or patterns within the molecules. Sun et al. used seven types of fingerprints as inputs to train the ML models and considered the influence of the fingerprint length on the prediction performance of different models to obtain diverse fingerprints. For instance, molecular access system (MACCS) fingerprints contained 166 bits and were the shortest input and the results were unsatisfactory due to their limited information.

Sun et al. showed the best combination of programming language and ML algorithm obtained using Hybridization fingerprints of 1024 bits and RF, to achieve a prediction accuracy of 81.76 percent; where Hybridization fingerprints represented SP2 hybridization states of molecules. When the fingerprint length increased from 166 to 1024 bits, the performance of all ML models improved since longer fingerprints included more chemical information.

To test the reliability of the ML models, Sun et al. synthesized 10 new OPV donor molecules. Then used three representative fingerprints to express the chemical structure of the new molecules and compared the results predicted by the RF model and the experimental PCE values. The system classified eight of the 10 molecules. The results indicated the potential of the synthetic materials for OPV applications with additional experimental optimization for two of the new materials. A minor change in structure could cause a large difference in PCE values. Encouragingly, the ML models identified such minor modifications to facilitate favorable prediction results.

In this way, Wenbo Sun and colleagues used a literature database on OPV donor materials and a variety of programming language expressions (images, ASCII strings, descriptors and molecular fingerprints) to build ML models and predict the corresponding OPV PCE class. The team demonstrated a scheme to design OPV donor materials using ML approaches and experimental analysis. They prescreened a large number of donor materials using the ML model to identify leading candidates for synthesis and further experiments. The new work can speed up new donor material design to accelerate the development of high PCE OPVs. The use of ML in conjunction with experiments will progress materials discovery.

The post Machine learning-assisted molecular design for high-performance organic photovoltaic materials appeared first on Artificial Intelligence.

No stranger to controversy, a Tony Stark reincarnate — Elon Musk — came out with an ominous prediction recently. “Forget North Korea, AI will start World War III” read the CNN headline. Elon Musk is not alone in fearing unintended consequences of the race to develop algorithms that we may or may not be able to control. Once a new technology is introduced it can’t be uninvented — Sam Harris points out in his viral TED talk. He argues that it’ll be impossible to halt the pace of progress, even if humankind could collectively make such a decision.

The critics and cheerleaders of AI alike agree on one thing: intelligence explosion will change the world beyond recognition.

While Bill Gates, Stephen Hawking and countless others are broadly on the same page with Musk and Harris, some of the leading thinkers recognize that AI, like any other technology, is value-neutral. Gunpowder, after all, was first used in fireworks.

Ray Kurzweil argues that “AI will be the pivotal technology in achieving [human] progress. We have a moral imperative to realize this promise while controlling the peril.” And, in his view, humanity has ample time to develop ethical guidelines and regulatory standards.

Making computers part of us, part of our bodies, is going to change our capabilities so much that one day, we will see our current selves as goldfish.

As the world edges towards singularity, future technology is bound to enhance the human experience in some way, and it is up to us to make sure it is for the better.

The critics and cheerleaders of AI alike agree on one thing: intelligence explosion will change the world beyond recognition. When thinking about the future, I found the metaphor offered by Vernor Vinge, on the Invisibilia podcast, especially stark: “making computers part of us, part of our bodies, is going to change our capabilities so much that one day, we will see our current selves as goldfish.” If this is a true extent of our expected AI-and-Tech-powered evolution, our contemporary norms and conventions go straight out of the window.

Putting the war and AI in the same sentence, we anthropomorphize the latter.

Even if the accurate predictions are a dud, shouldn’t we at least attempt to apply the prism of exponential technologies to review our basic assumptions, question fundamentals of human behavior, and scrutinize our societal organization? AI’s promise could be an apocalypse or eternal bliss or anything in between, but, as we speculate on the outcome, we are making a value judgment. And here we ought to recognize our susceptibility to the projection bias. It compels us to apply the present-day intellectual framing to ponder the future.

Putting the war and AI in the same sentence, we anthropomorphize the latter. When we worry about the robots and machine-intelligence causing mass unemployment, we must recognize that such anxiety is only justified if human labor remains an economic necessity. When we say that the spiraling-out-of-control tech progress will create more inequality, we assume that the idea of private property, wealth, and money will survive the fourth-industrial revolution.

It’s an arduous task to define the fundamental terms, much less to question them. But, perhaps, playing out a couple of scenarios could prove a useful exercise is circumventing projection bias.

Competition & Collaboration

The natural selection is, at its core, a multidimensional competition of traits and behaviors. It manifests itself in a basic competitive instinct that humans are all too familiar with. Evolutionary psychology postulates that the driver of human behavior is a need to perpetuate one’s genes. So Homo Sapiens evolved competing for mates, fighting for resources to feed the offspring, all with a singular objective to maximize their genes’ chances to be passed on.

When the algorithms are better at decision-making than humans, and we surrender much of our autonomy to them, how will our competitive instinct fare?

On the other hand, we are, according to Edward O. Wilson, “one of only two dozen or so animal lines ever to evolve eusociality, the next major level of biological organization above the organismic. There, group members across two or more generations stay together, cooperate, care for the young, and divide labor…” In other words, we might have to attribute the stunning success of our species to the fine balance we’ve maintained between competition and cooperation instincts.

What will be the point of resource competition in the world of abundance?

Whether general machine intelligence is imminent or even achievable, the idea of post-scarcity economy is gaining ground. If and when the automation of pretty much everything delivers the world where human labor is redundant, what will be the wider ramifications for our value system and societal organization? When the algorithms are better at decision-makingthan humans, and we surrender much of our autonomy to them, how will our competitive instinct fare?

What will be the point of resource competition in the world of abundance? Is it possible that our instinct to compete slowly evaporates as a useful construct? Could we evolve to live without it? Unlike ants and bees that cooperate on the basis of rigid protocols, humans are spectacularly adaptable in our cooperation abilities. According to Yuval Harari, that’s what ultimately underpinned the rise of sapiens to dominate the Earth. Is it conceivable that the need to compete turns into an atavism as the technic transformations described by Kurzweil begin to materialize?

Economy

How can we be sure that the basic pillars of our economic thinking (e.g. private property, ownership, capital, wealth, etc.) will survive post-scarcity? 100 years from now, will anybody care for labor productivity? How relevant could our policies encouraging employment be when all of the humanity is freeriding on the “efforts” of the machines? What are we left with, when the basics such as supply and demand have been shuddered?

If ownership is pointless and money is no longer a useful unit of exchange, how will we define status?

To a gainfully employed person, today, a prospect of indefinite leisure might appear more of a curse than a blessing. This sentiment, viewed through the lens of natural selection, makes sense. The economic contribution by all able members of society would’ve been preferred to the mass pursuit of idleness. But should we be projecting the same trend into the future? What may sound like decadence and decay to us now, may be construed quite differently in the world no longer powered by the known economic forces.

The working assumption is that no matter what, someone will have to own the machines and pay for goods and services. Yet the idea of property and money is nothing more than social constructs that we all agreed on. If ownership is pointless and money is no longer a useful unit of exchange, how will we define status?

The post Artificial Intelligence: The Good, The Bad, and The Unfathomable appeared first on Artificial Intelligence.