Cloud Hosting

image:INL researchers perform experiments using the Temporal Analysis of Products (TAP). view more

Credit: Idaho National Laboratory

If you follow the directionsin a cake recipe, you expect to end up with a nice fluffy cake.In Idaho Falls,though, the elevation can affecttheseresults.When baked goods dont turn outas expected, the troubleshooting begins.This happens in chemistry,too.Chemistsmustbeable to account for how subtle changes or additions may affect the outcome for better or worse.

Chemists maketheir version ofrecipes, known as reactions,to create specific materials.These materialsare essential ingredients to an array of products found in healthcare, farming, vehicles andother everyday productsfrom diapers to diesel.When chemists develop new materials, they rely on information from previous experiments and predictions based onpriorknowledge ofhowdifferent starting materials interact with others and behave underspecificconditions.There are a lot of assumptions, guesswork and experimentation in designing reactions using traditional methods.New computational methods like machine learning can help scientists better understand complex processes like chemical reactions.While it can be challenging forhumans topick outpatternshiddenwithin the data from many different experiments, computers excel at this task.

Machine learning isan advancedcomputational toolwhereprogrammers givecomputerslots ofdata andminimalinstructions about how to interpret it. Instead of incorporatinghuman bias into the analysis, the computer isonly instructed to pull out what it finds to be important from the data. This could be an image of a cat (if the input is all the photos on the internet) orinformation about how a chemical reactionproceeds through a series ofsteps, as is thecasefora set of machine learning experiments that are ongoing at Idaho National Laboratory.

At the lab,researchersworking with the innovative Temporal Analysis of Products (TAP)reactorsystemaretryingto improveunderstanding of chemical reactions by studying the role of catalysts,whicharecomponentsthat can be added toamixture of chemicals to alter thereactionprocess.Oftencatalystsspeed up thereaction,but they can do other things,too. In baking and brewing,enzymesact as catalyststo speed up fermentationandbreakdown sugars in wheat (glucose) into alcohol and carbon dioxide,which creates the bubbles that make bread riseand beer foam.

In the laboratory,perfectinga new catalystcan be expensive, time-consuming and even dangerous.According toINLresearcher Ross Kunz, Understanding how and why a specific catalyst behavesin a reaction is theholygrail ofreaction chemistry.To help find it,scientists arecombiningmachine learningwith a wealth of new sensor datafrom the TAP reactorsystem.

The TAP reactor system uses an array of microsensors to examine the different componentsof a reaction in realtime.For the simplestcatalytic reaction,the system captures8uniquemeasurementsin each of 5,000timepointsthat make up the experiment.Assembling the timepoints into a single data set provides 165,000 measurements foroneexperiment on a very simple catalyst.Scientiststhenuse the datatopredict what is happening in the reaction at a specific timeand how different reaction steps work together in a larger chemical reaction network.Traditional analysis methods canbarelyscratch the surfaceofsuch a large quantity of datafor a simple catalyst, let alonethe many more measurements thatare produced by acomplex one.

Machine learning methods can take theTAP dataanalysis further. Using a type of machine learning called explainableartificial intelligence, orAI,theteam caneducatethe computer about known properties of thereactionsstarting materialsand the physics that govern these types of reactions, a process called training.The computer can apply thistrainingand the patterns that it detects in the experimental data to better describe theconditions inareactionacross time.The team hopes that theexplainable AI method will produce adescription of the reaction that can be used toaccuratelymodelthe processes that occur during theTAP experiment.

In most AI experiments, a computer is given almost no trainingon the physicsand simply detects patterns in the data based upon what it can identify,similar tohow a baby might react to seeing something completely new.By contrast,the value of explainable AI lies in the fact that humanscan understand the assumptions and information that lead to the computers conclusions.This human-level understanding can make it easier for scientists to verify predictions and detect flaws and biases in the reaction description produced by explainable AI.

Implementing explainable AIis not as simple or straightforward as it might sound.With support from the Department of Energys Advanced Manufacturing office, theINLteam has spent two years preparing theTAPdata for machine learning,developing andimplementingthe machinelearning program, andvalidating the results for a common catalyst in a simple reaction that occursinthe car you driveeveryday. This reaction,the transformation of carbon monoxideinto carbon dioxide,occurs ina carscatalytic converter andrelies onplatinumasthe catalyst. Since this reaction is well studied,researcherscan checkhow well the results of the explainable AI experiments match known observations.

In April 2021, the INL team published their results validating the explainable AI method with the platinum catalyst in the article Data driven reaction mechanism estimation via transient kinetics and machine learninginChemical Engineering Journal.Now that the team has validated the approach, they are examining TAP data frommore complex industrialcatalystsused in the manufacture of smallmolecules like ethylene, propylene and ammonia. They are also working with collaborators at Georgia Institute of Technologyto applythemathematical models that result from themachine learningexperiments tocomputersimulationscalled digital twins. This type of simulation allows the scientists topredict what will happen if they change an aspectof the reaction. When a digital twin is based on avery accurate model of a reaction, researcherscanbe confident in itspredictions.

Bygivingthe digital twinthe taskto simulate a modification to a reaction or new type of catalyst, researchers can avoid doing physical experiments for modifications that are likely to lead to poor results or unsafe conditions. Instead,the digital twin simulation can savetime and moneyby testing thousands of conditions,while researchers can testonly a handful of the mostpromising conditions in the physical laboratory.

Plus, this machine learning approach can produce newer and more accurate modelsfor each new catalyst and reaction condition testedwith the TAP reactorsystem.In turn, applying these models to digital twin simulations gives researchers the predictive power to pick the best catalysts and conditions to test next in the TAP reaction. As a result, each roundof testing, model development and simulationproducesa greater understanding of how a reactionworksand howtoimprove it.

These toolsarethe foundation of a new paradigm incatalyst science but alsopave the way for radical new approaches inchemical manufacturing,said Rebecca Fushimi, who leads the project team.

About Idaho National LaboratoryBattelle Energy Alliance manages INL for the U.S. Department of Energys Office of Nuclear Energy. INL is the nations center for nuclear energy research and development,and alsoperforms research in each of DOEs strategic goal areas: energy, national security, science and the environment. For more information, visitwww.inl.gov. Follow us on social media:Twitter,Facebook,InstagramandLinkedIn.

Chemical Engineering Journal

Data driven reaction mechanism estimation via transient kinetics and machine learning

18-Apr-2021

Disclaimer: AAAS and EurekAlert! are not responsible for the accuracy of news releases posted to EurekAlert! by contributing institutions or for the use of any information through the EurekAlert system.

Read the original here:
Learning to improve chemical reactions with artificial intelligence - EurekAlert

AI in telecom

The number of smartphone users across the world has skyrocketed over the last decade and promises to do so in the future too. Additionally, most business functions can now be executed on mobile devices. However, despite the mobile surge, telecom operators around the world are still not that profitable, with average net profit margins hovering around the 17% mark. The main reasons for the middling profit rates are the high number of market rivals vouching for the same customer base and the high overhead expenses associated with the sector. Communication Service Providers (CSPs) need to become more data-driven to reduce such costs and, automatically, improve their profit margins. Increasing the involvement of AI in telecom operations enables telecom companies to make this switch from rigid, infrastructure-driven operations to a data-driven approach seamlessly.

The inclusion of AI in telecom functional areas positively impacts the bottom line of CSPs in several ways. Businesses can use specific capabilities, avatars or applications of machine learning and AI for this purpose.

Mobile networks are one of the prime components of the ever-expanding internet community. As stated earlier, a large number of internet users and business operations have gone mobile in recent times. Additionally, the emergence of 5G and edge applications, and the impending arrival of the metaverse, will simply increase the need for high-performance telecom networks. It is very likely that the standard automation tech and personnel will be overwhelmed by the relentless pressure of high-speed network connectivity and mobile calls.

The use of AI in telecom operations can transform an underperforming mobile network into a self-optimizing network (SON). Telecom businesses can monitor network equipment and anticipate equipment failure with AI-powered predictive analysis. Additionally, AI-based tools allow CSPs to keep network quality consistently high by monitoring key performance indicators such as traffic on a zone-to-zone basis. Apart from monitoring the performance of equipment, machine learning algorithms can also continually run pattern recognition while scanning network data to detect anomalies. Then, AI-based systems can either perform remedial actions or notify the network administrator and engineers in the region where the anomaly was detected. This enables telecom companies to fix network issues at source before they adversely impact customers.

Network security is another area of focus for telecom operators. Of late, the rising security issues in telecom networks have been a point of concern for CSPs globally. AI-based data security tools allow telecom companies to constantly monitor the cyber health of their networks. Machine learning algorithms perform analysis of global data networks and past security incidents to make key predictions of existing network vulnerabilities. In other words, AI-based network security tools enable telecom businesses to pre-empt future security complications and proactively take preventive measures to deal with them.

Ultimately, AI improves telecom networks in multiple ways. By improving the performance, anomaly detection and security of CSP networks, machine learning algorithms can enhance the user experience for telecom company clients. This will result in a growth of such companies customer base in the long term, and, by extension, an increase in profits.

How Telecom Companies Can Leverage Machine Learning To Boost Their Profits

The Europol classifies the telecom sector to be particularly vulnerable to fraud. Telecom fraud involves the abuse of telecommunications systems such as mobile phones and tablets by criminals to siphon money off CSPs. As per a recent study, telecom fraud accounted for losses of US$40.1 billionapproximately 1.88% of the total revenue of telecom operators. One of the common types of telecom fraud is International Revenue Sharing Fraud (IRSF). IRSF involves criminals linking up with International Premium Rate Number (IPRN) providers to illegally acquire money from telecom companies by using bots to make an absurdly high number of international calls of long duration. Such calls are difficult to trace. Additionally, telecom companies cannot bill clients for such premium calls as the connections are fraudulent. So, telecom operators end up bearing the losses for such calls. The IPRNs and criminals share the spoils between themselves. Apart from IRSF, vishing (a portmanteau for voice calls and phishing attacks) is a way in which malicious entities dupe clients of telecom companies to extract money and data. The involvement of AI in telecom operations enables CSPs to detect and eliminate these kinds of fraud.

Machine learning algorithms assist telecom network engineers with detecting instances of illegal access, fake caller profiles and cloning. To achieve this, the algorithms perform behavioral monitoring of the global telecom networks of CSPs. Accordingly, the network traffic along such networks is closely monitored. The pattern recognition capabilities of AI algorithms come into play again as they enable network administrators to identify contentious scenarios such as several calls being made from a fraudulent number, or blank callsa general indicator of vishingbeing repeatedly made from questionable sources. One of the more prominent examples of telecom companies using data analytics for fraud detection and prevention is Vodafones partnership with Argyle Data. The data science-based firm analyzes the network traffic of the telecom giant for intelligent, data-driven fraud management.

Detecting and eliminating telecom fraud are major steps towards increasing the profit margins of CSPs. As you can see, the role of AI in telecom operations is significant for achieving this objective.

To reliably serve millions of clients, telecom companies need to have a massive workforce that can handle their backend operations on a daily basis efficiently. Dealing with such a large volume of customers creates several opportunities for human error.

Telecom companies can employ cognitive computinga robotics-based field that involves Natural Language Processing (NLP), Robotic Process Automation (RPA) and rule enginesto automate the rule-based processes such as sending marketing emails, autocompleting e-forms, recording data and carrying out certain tasks that can replicate human actions. The use of AI in telecom operations brings greater accuracy in back-office operations. As per a study conducted by Deloitte, several executives in the telecom, media and tech industry felt that the use of cognitive computing for backend operations brought substantial and transformative benefits to their respective businesses.

Customer sentiment analysis involves a set of data classification and analysis tasks carried out to understand the pulse of customers. This allows telecom companies to evaluate whether their clients like or dislike their services based on raw emotions. Marketers can use NLP and AI to sense the "mood" of their customers from their texts, emails or social media posts bearing a telecom companys name. Aspect-based sentiment analytics highlight the exact service areas in which customers have problems. For example, if a customer is upset about the number of calls getting dropped regularly and writes a long and incoherent email to a telcos customer service team about it, the machine learning algorithms employed for sentiment analysis can still autonomously ascertain their mood (angry) and the problem (the call drop rate).

Apart from sentiment analysis, telecom businesses can hugely benefit from the growing emergence of chatbots and virtual assistants. Service requests for network set-ups, installation, troubleshooting and maintenance-based issues can be resolved through such machine learning-based tools and applications. Virtual assistants enable CRM teams in telecom companies to manage a large number of customers with ease. In this way, CSPs can manage customer service and sentiment analysis successfully.

Across the board, users generally cite the quality of their telecom customer service to be below satisfactory. Telecom users are constantly infuriated by long waiting times to get to a service executive, unanswered complaint emails and poor grievance handling by CSPs. Poor CRM does not bode well for telecom companies as it maligns their reputation and diminishes shareholder confidence. By implementing machine learning for CRM, telecom companies can address such issues efficiently.

Like businesses in any other sector, telecom companies need to boost their profits for long-term survival and diversification. As stated at the beginning, there are multiple factors that thwart their chances of profit generation. Going down the data science route is one of the novel ways to overcome such challenges. By involving AI in telecom operations, CSPs can manage their data wisely and channelize their resources towards maximizing revenues.

Despite the positives associated with AI, only a limited percentage of telecom businesses have incorporated the technology for profit maximization. Gradually, one can expect that percentage to rise.

More:
How Telecom Companies Can Leverage Machine Learning To Boost Their Profits - Forbes

Data collection

To prepare the training data and labels, we downloaded the whole-genome sequencing (WGS) data for 10,575 MTB isolates from the sequence read archive (SRA) database17 and obtained corresponding lineage and phenotypic drug susceptibility test (DST) data from CRyPTIC Consortium and the 100,000 Genomes project in an excel file, which is also available in the supplementary of their publication15. The phenotypic DST results for the drugs were used as labels when training and evaluating our ML models. All the data were collected and shared by the CRyPTIC Consortium and the 100,000 Genomes Project15. Like the datasets used by previous studies, this dataset is imbalanced in that most isolates are susceptible, and the minority of them are resistant for all the four first-line drugs (Fig.1) and four second-line drugs. The numbers of isolate samples with phenotypic DST results available are 7138, 7137, 6347 and 7081 for EMB, INH, PZA and RIF, respectively. There are 6291 shared isolates among the four sample sets. In addition, 6820 out of the 10,575 isolates have phenotypic DST result available for each of the four second-line drugs.

Phenotypic overview of the MTB isolates. This bar chart shows numbers of susceptible and resistant isolates with DST results available for each of the four first-line drugs.

To detect the potential genetic features that could contribute to MTB drug resistance classification, we used a command-line tool called ARIBA18. ARIBA is a very rapid, flexible and accurate AMR genotyping tool that generates detailed and customizable outputs from which we extracted genetic features. First, we downloaded all reference data from CARD, which included not only references from different MTB strains but also from other bacteria (e.g., Staphylococcus aureus). Secondly, we clustered reference sequences based on their similarity. Then we used this collection of reference clusters as our pan-genome reference and aligned read pairs of an isolate to them. For each cluster that had reads mapped, we ran local assemblies, found the closest reference, and identified variants. After running these steps, ARIBA generated files including a summary file for alignment quality, a report file containing information of detected variants and AMR-associated genes, and a read depth file. For each cluster, the read depth file provides counts of the four DNA bases on each locus of the closest reference where reads were mapped.

Next, we filtered out low-quality mappings that did not pass the match criteria defined in ARIBAs GitHub wiki18. From these high-quality mappings, we collected novel variants in coding regions, well-studied resistance-causing variants and AMR-associated gene presences that were detected from at least one out of the 10,575 isolates as 263 genetic features. In addition, we included indicator variables for each of the 19 lineages into our feature vector resulting in a total of 282 features.

We applied two traditional ML algorithms, RF and LR, on the sample sets labeled with phenotypic DST results (see Data collection section) to train MTB AMR classifiers for the eight drugs (first-line and second-line), where the feature vector for each sample consists of the 282 features mentioned in Genetic feature extraction section.

RF is an ensemble method and made up of tens or hundreds of estimators (decision trees) to compress overfitting19,20. A final prediction is an average or majority vote of the predictions of all trees. It is often used when there are large training datasets and a large number of input features. Moreover, RF is good at dealing with imbalanced data by using class weighting. Here we trained each RF classifier with 1000 estimators.

LR is a popular regression technique for modeling binary dependent variable21. By using a sigmoid function (logit), linear regression is transformed into logistic regression so that the prediction range is [0, 1] for outputting probabilities. Then, LR model is fitted using maximum likelihood estimation. During the training process, we applied L1 regularization on LR models for feature selection and to prevent overfitting22.

CNN is a class of deep neural networks that takes multi-dimensional data as input23. When we sayCNN, generally, we refer to a 2-dimensional CNN, which is often used for image classification. However, there are two other types of CNN used in practice: 1-dimensional and 3-dimensional CNNs. Conv1D is generally used for time-series data where the kernel moves on one dimension and the input and output data are 2-dimensional. Conv2d and 3D kernels move on two dimensions and three dimensions, respectively.

Because deep learning algorithms require substantial computational power, we performed feature selection to only keep relevant features as input for deep learning algorithms. First, we randomly selected 80 percent of samples to calculate the importance of each feature by using the scikit-learn RF feature importance function that averages the impurity decrease from each feature across the trees to determine the final importance of each variable24. Then, we tuned the feature importance cutoff to find the one that maximizes the F1-score of an RF model trained on the remaining 20 percent of samples. For each of the eight drugs, features were selected when their feature importance scores were bigger than the optimal cutoff. The tuning processes for first-line drugs are visualized in Fig.2.

Feature importance cutoff tuning. For the four first-line drugs, when the cutoff increases, the F1-score quickly increases to its maximum and then continues to decrease. The cutoffs maximized F1-scores are 0.0004 (EMB), 0.0006 (INH), 0.0008 (PZA) and 0.0016 (RIF).

After the relevant features were selected, we designed and built a multi-input CNN architecture with TensorFlow Keras25 that took N inputs of 421 matrices representing N selected SNP features into the first layer. Each 421 matrix consists of normalized DNA base counts for each locus within a 21-base reference sequence window centered on the focal SNP (Fig.3). We generated normalized counts based on the raw base counts extracted from the read depth file mentioned in Genetic feature extraction section. Our convolutional architecture starts with two 1D convolutional layers followed by a flattening layer for each SNP input. Then, it concatenates the N flattening layers with the inputs of AMR-associated gene presence and lineage features. Finally, we added three fully connected layers to complete the deep neural network architecture (Fig.4). It smoothly integrates sequential and non-sequential features.

Conversion of raw base counts at each locus of a 21-base reference window into normalized base counts as Conv1D input of each selected SNP feature. The raw base counts were derived from reference-reads alignment, as shown on the left of this figure. The center of the window is the locus of a selected SNP feature. The normalized base counts at each locus are the percentage of the four DNA bases (ACGT), respectively.

Flowchart of our 1D CNN architecture.

Read more:
Accurate and rapid prediction of tuberculosis drug resistance from genome sequence data using traditional machine learning algorithms and CNN |...