Technology Watch AI and Chemical Industry€¦ · •Information retrieval •What AI can’t do… ( very well ) AI approach AI opportunities in chemistry today Patent landscape

CONFIDENTIAL

Ref: DOCLOG-XXXX-DOC-A (edit in slide master)Document Title - yyyy.mm.dd (edit in slide master) 1

TEM

P-00

10-D

OT-

F -V

erha

ert P

rese

ntat

ion

Technology Watch AI in Chemical Industry

SMART INDUSTRY

Jochem Grietens

Applied physics engineer – AI engineer

[email protected]

13.09.2019

CONFIDENTIAL

AI technology review – 13.09.2019 2

Content

Demystification• What is AI ?• What is AI good at ?

• Classification • Finding patterns • Recognizing deviations from patterns• Predicting• Structuring the unstructured• Estimating from proxy information• Agency• Model complex systems• Optimization, search, planning• Information retrieval

• What AI can’t do… ( very well )AI approachAI opportunities in chemistry todayPatent landscape

CONFIDENTIAL


CONFIDENTIAL

DemystificationWhat is AI ?What is AI good at ?What AI can’t do… ( very well )

CONFIDENTIAL


CONFIDENTIAL

What is AI ?

CONFIDENTIAL


What is AI ?

There is a lot of debate outside of the A.I. community on how to define the field. The experts more or less agree:

Artificial intelligence“The theory and development of computer systems able to perform cognitive tasks normally requiring human intelligence. “

Cognitive tasks are defined in cognitive science as : attention/logic and reasoning, decision making, perception ( such as visual perception, speech/sound recognition), language understanding and generation (translation).

CONFIDENTIAL


AI effect

The above definition leads to problems because of the AI effect:

“As machines become increasingly capable, tasks considered to require "intelligence" are often removed from the definition of AI, a phenomenon known as the AI effect. A quip in Tesler's Theorem says "AI is whatever hasn't been done yet”. For instance, optical character recognition is frequently excluded from things considered to be AI, having become a routine technology.”

The field is moving on different fronts but the main advances have been in prediction and perception and were made possible by deep learning: speech recognition, visual perception, …

https://en.wikipedia.org/wiki/AI_effect

https://en.wikipedia.org/wiki/Optical_character_recognition

CONFIDENTIAL


What about General vs Narrow AI

Long term aimDevelop systems that achieve a level of cognitive performance similar/comparable/better than that of humans. (General AI)

Not in the near future, no practical or even noteworthy academic systems exist now with such capabilities. Should not be the focus of companies today.

Short term aimOn specific tasks that seem to require intelligence: Develop systems that achieve a level of cognitive performance similar/comparable/better than that of humans. (Narrow AI)

Achieved for many tasks already, can speed up your business today. Field is moving really quickly.

CONFIDENTIAL


AI Definition

To achieve flight, humans did not have to imitate birds exactly.

• The principles of flight were extracted ( lift surface + velocity)

• The EFFECT of flight was achieved.

At the very least in the context of the short term aim of AI:

• we do not want to imitate human intelligence.

• Reproduce the EFFECT of intelligence

CONFIDENTIAL


The field of data science (non-exhaustive)

Data science

AI

Classical AI techniques

Search / Planning

Optimization

Logic: induction and deduction

Knowledge representation

Expert systems

ML

Supervised machine learning

Bayesian networks

Decision trees, SVM’s, …

Neural networks

Unsupervised machine learning

Clustering

Reinforcement learning

Learning distributions: autoencoders, GANS, …

PCA

…Data analytics

Statistics

Machine learning

Clustering

...

Mathematics

The field of AI draws from many fields of study, in this tree a non-exhaustive overview is given in an attempt to provide context. The relations are explained further in the slides below.

CONFIDENTIAL


The field of data science

Data science is a multi-disciplinary field that uses scientific methods, processes, algorithms and systems to extract knowledge and insights from structured and unstructured data. The data science field entails data analytics and A.I. among others. A useful distinction can be made by observing:

• Data analytics is a field where a data scientist stays in the driver seat to extract information, draw conclusions and answer questions.

• A.I. systems generally require an expert for development but afterwards they can perform the required tasks in an automated way.

However, data analysts use AI/ML techniques (clustering, classifying) and visa versa.

CONFIDENTIAL


Classical subfields of AI

The “classical field” of AI entails • Search/planning/scheduling• Optimization • Logic induction and deduction• Knowledge representation, expert systems, …

Although these fields are less novel, they are still highly relevant to solve contemporary problems because:

• Progress is still being made on the scientific front• Through the availability of data and computational power, problems that were

previously not solvable have now become candidates to apply these techniques

• Advances in machine learning allow for unstructured data such as speech, text, images etc. to be interpreted and translated to structured data. This structured data can be handled by these classical subfield. This creates tremendous synergy.

CONFIDENTIAL


Machine learning

“Machine learning is a branch of artificial intelligence that uses sophisticated algorithms to give computers the ability to learn from the data and make predictions.”

The biggest leaps in AI of the last decade have been in the machine learning space. These advances have been made possible by 3 main factors in order of decreasing importance: • Computing power ( parallel computing made

Available/affordable through the gaming world)• Advances in the ML techniques. • Data availability

The momentum created by the interest of the general public, large companies and governments has greatly contributed to funds and efforts being directed towards A.I. This has created the perfect storm and the avalanche of innovation we currently observe.

CONFIDENTIAL


Machine learning steps

Machine learning comprises 2 steps: • During development – Training. Data is fed to the ML algorithm, the

algorithm learns patterns from the given examples. • During operation – Inference. The model is deployed and in operation,

new data is fed to the model and the learned patterns can be applied to new input data to provide the desired output.

CONFIDENTIAL


Machine learning FAQ

Frequently asked question : “ But don’t machine learners learn continuously ? “ Answer: In most applications, the two steps of machine learning are clearly separated in time. Training is performed and a fully trained network is deployed. But, • These steps are often repeated iteratively and

alternately on new batches/instances of data. This called iterative learning and allows for incremental improving and releasing of new models.

• Models that learn with every new incoming data point exist as well. The inference step and training step happen simultaneously. This is called continuous learning. These techniques are not widespread in engineering applications with high reliability requirements yet, because they are harder to test, verify and validate before release, since they are ever-changing.

CONFIDENTIAL


Supervised vs. Unsupervised learning

Supervised learning algorithms require annotated training data containing both:• Example input data• Associated desired output data.

The models then learns to extract the desired output form the input data.

CONFIDENTIAL


Supervised vs. Unsupervised learning

Unsupervised learning algorithms require training data containing only:• Example input data

The models extract patterns from the input data and apply these to new data to provide insight.

CONFIDENTIAL


Machine learning FAQ

Frequently asked question : “ Does this mean that unsupervised machine learning algorithms are smarter and better than supervised machine learning algorithms ? Since they have no need for annotated data ? “ Answer: No, these types of models are used for different tasks and have different characteristics. Unsupervised models are not able to perform many of the tasks supervised machine learning algorithms do very well and visa-versa.

CONFIDENTIAL


Machine learning overview

There are many machine learning paradigms and algorithms. These are some of the more important families of models: Bayesian (belief) networks, Support vector machines, Decision trees/forests, Artificial neural networks and many more…

All these families have their specific characteristics.

• Amount of data required

• Data noise sensitivity

• Computational effort required for training and inference.

• Human interpretability of the learned patterns: Black box vs. White box

• Performance

• Supervised vs. supervised.

Choosing the right ML for the job should be based on requirements.

CONFIDENTIAL


Machine learning model selection FAQ

QUESTION: “ All the material I read about AI talks about neural networks, are they the best overall models out there right now ? ” Yes, and no. The main driver behind the new wave of AI technologies has been neural networks. The main reason is that these networks turn out to be remarkably versatile in several regards: • The types of tasks they can solve ( estimation, language modeling, speech-to-text, prediction,

computer vision, … )• The complexity of relations they can learn. (simple to highly complex)• The amount of data they can handle and learn from. (from small data to big data)• Their robustness to noise in the data.

Because of this flexibility and these models have taken the AI world by storm. However, A good selection should match the AI task requirements and model characteristics. Although neural networks have achieved exceptional results and have facilitated the revival of AI, they have some drawbacks regarding interpretability and computational cost. In specific cases these drawbacks might lead the developer to favor other ML techniques. These limitations should be well understood. That being said, NN have revolutionized the AI world and the rate of innovation is increasing in speed partly because of them.

CONFIDENTIAL


AI = Big data FAQ

Question: “ Is A.I. inseparably tied to BIG data or does A.I. for small data exist ? “ Answer: No it is not, however it is often desired. Let’s elaborate,1. Firstly, not all A.I. techniques are data driven. A lot of search, planning and

optimization methods just require a good description of the problem. 2. Secondly, even some classes of machine learning methods can perform well on

limited amount of data, given a limited complexity of the task.3. Thirdly, many of the very high performance ML techniques for complex tasks do

require big data. i.e. large, deep neural networks require a lot of data. However, for common tasks such object recognition we can reuse networks that were trained on other dataset and only have be fine-tuned on reduced dataset that is specific to our problem. This technique is called transfer learning.

To conclude, AI requires big data for complex problems of uncommon tasks that are very specific to your use case. Complexity is dependent on the amount of input and output variables and the complexity of the relations that needs to be learned.

CONFIDENTIAL


CONFIDENTIAL

What AI is good at… • Classification• Predictions• Recognize patterns• Recognize deviation from patterns• Structuring the unstructured• Estimating from proxy information• Agency• Model complex systems• Optimization, search

CONFIDENTIAL


CONFIDENTIAL

Classification

CONFIDENTIAL


Classifiers learn to classify samples based on their features. • The input data can take any data format: images, videos, text,

molecule representations, … • The output is a finite set of classes that we want to recognize.

ML based classification models can achieve fully automated above human performance in many cases.

Classification

CONFIDENTIAL


Classification – example use cases

• Machine Learning Based Toxicity Prediction. From Chemical Structural Description to toxicity classification.

• Computer-Aided drug design. 743,336 compounds, approximately 13 million chemical features, and 5069 drug targets were used to train the ML algorithm. The model provides classification of properties, structures and functions.

CONFIDENTIAL


CONFIDENTIAL

Finding patterns

CONFIDENTIAL


Find patterns

ML algorithms are good at finding patterns in data of any type. The power of these algorithms becomes apparent when the data is to large for humans to sift true.

Examples of every day pattern finding powered by ML: • Spam filters find patterns in spam

mails to later classify and exclude them

• Recommender engines find patterns in consumers profiles and products to match sales and allow targeted advertising.

• Time series patterns allow to predict stock prices.

• Clustering algorithms to find similar compounds to a target chemical compound.

CONFIDENTIAL


Find patterns – example use cases

• Deep Reinforcement Learning Approaches for Process Control . Finding patterns in plant behavior for process control.

• Sales Lead Scoring decision support -Pattern finding ML systems allow to learn from CRM historical data to find companies with a high chance of closing. These tools extract patterns of previously successful sales pipelines and search for companies with similar patterns.

• Pattern finding ML systems allow to learn from historical crm data to find customer with high up-sell or cross-sell potential. These tools extract patterns of previously successful sales pipelines and search for companies with similar patterns.

CONFIDENTIAL


CONFIDENTIAL

Recognize deviation from patterns

CONFIDENTIAL


Recognize deviation from patterns

Oftentimes we want to recognize deviations from ‘normal operation’ of a system. These deviations might be very rare or no data is available of them at all.

Pattern recognition systems rely on many similar examples of these patterns being available to learn from. This approach won’t work for detecting deviations from patterns.

In general these cases are solved by using ML techniques characterizing normal behavior and detecting when the system deviates from this behavior.

Using ML for this purpose allows to characterize highly complex systems behavior and predicting never seen before anomalies.

CONFIDENTIAL


Recognize deviation from patterns – example use cases

• On-line reactor monitoring with neural networks. On-line condition monitoring and signal validation has become a significant issue to ensure stable operation and deviations from normal operations produce alerts.

• Deep learning for pyrolysis reactor monitoring. From thermal imaging toward smart monitoring system to detect faults using neural networks.

• Historical example: This example from Suewatanakul [1993] demonstrates the use of a feedforward ANN to detect faults in a heat exchanger.

CONFIDENTIAL


CONFIDENTIAL

Prediction

CONFIDENTIAL


Predicting

ML algorithms learn how to predict desired output parameters from new, never seen before samples. The ML learns from record data or historical data.

Predictions can predict quantities real time or predict into the future.

The input data can contain multiple variables taking into account many context variables other methods would not be able to handle.

CONFIDENTIAL


Predicting – Example use cases (1)

• Sales forecasting - Sales managers face the daunting challenge of trying to predict where their team’s total sales numbers will fall each quarter. Using an AI algorithm, managers are now able to predict with a high degree of accuracy next quarter’s revenue.

• Quantum chemistry - Machine learning algorithm to predict the atomization energies of organic molecules. (von Lilienfeld)

• Computational Material Design – ML ( deep learning ) applications to predict and design material properties in silico.

• Thermal reactor control - High-speed and high-accuracy thermal control of a continuous-flow chemical reactor with computer vision and a predictive Artificial Neural Network.

CONFIDENTIAL


Predicting – Example use cases (2)

• Chemical reaction prediction - Chemists at Princeton University and Spencer Dreher of Merck Research Laboratories harness artificial intelligence to predict the future of chemical reactions. They predict yields accurately while varying up to four reaction components by applying machine learning.

• Chemical reaction prediction - treating chemical reactions as a translation problem ( think google translate) . In using such an approach, researchers were able to feed chemical components into a neural network trained on a dataset of 395,496 reactions. The neural network then used what it had learned about prior reactions to make predictions about what would occur under new conditions.

• Predictive maintenance - Predictive maintenance is the practice of using anomaly detection, pattern recognition and other AI techniques to predict when machinery needs maintanence. This is being applied in factories, fleet management, process control today.

CONFIDENTIAL


CONFIDENTIAL

Structuring the unstructured

CONFIDENTIAL


Structuring the unstructured

For the longest time, computers could only perform operations on structured data like excel sheets, databases etc. Advances in neural networks have revolutionized computing by allowing unstructured data to be interpreted and structured in meaningful ways. This has allowed unstructured formats such as natural language ( written and spoken), images, videos, speech and others to be converted to structured data by means of extracting higher level meaning and features from those documents. These advances have added perception to computers resulting in an explosion of applications that used to be off-limits for computers.

Examples from daily life: • Adding perception systems to cars enabling autonomous cars. • Computer vision: object detection, face recognition and others for identification. • Speech to text and natural language understanding allowing for voice interfaces to computers• …

Semantic segmentation = automatically assigning a meaningful label to each pixel.

CONFIDENTIAL


Structuring the unstructured – example use cases (1)

• Computer vision enabled techniques for organic synthesis.

• Lab tools – Computer vision and speech recognition technologies for experiment tracking, monitoring and logging.

• Semantic segmentation on molecules - multi-scale structural analysis of proteins by deep semantic segmentation

• Deep learning to yield a powerful tool for both protein design and structure prediction.

CONFIDENTIAL


Structuring the unstructured – example use cases (2)

• Speech to text and language understanding technologies allow interfacing with devices in new, hands-free ways.

• Production line intelligence -Rockwell automation created a ‘data scientist in a box ‘ called Shelby. Including a production line chatbot with text based conversational interface, chatbot and a voice interface. Based on Microsoft Cortana.

http://www.ssi-schaefer-asia.com/blogsg/automation/facing-up-to-asias-manpower-challenge/

http://www.ssi-schaefer-asia.com/blogsg/automation/facing-up-to-asias-manpower-challenge/

https://www.bc.edu/cleanroom

https://www.bc.edu/cleanroom

CONFIDENTIAL


CONFIDENTIAL

Estimating from proxy information

CONFIDENTIAL


Estimate from proxy information

Estimators to estimate the unmeasured quantities indirectly by using proxy-parameters of measured quantities. The machine learning algorithms then learns the relation between the measured parameters and the desired unmeasured parameters.

This is often useful because some quantity can not be measured directly, so it needs to be estimated from related parameters that can be measures.

Example:• Extracting the letters you intended to type on your smartphone keyboard from the

letters you actually typed ( autocorrect ) • Predictive maintenance by measuring vibrations of an accelerometer on a machine

to detect mechanical failure. • Estimating core temperature from multiple external temperatures• Estimating process quantity in a reactor vessel that is too hot for direct

measurement but has some surrounding parameters that are linked to the condition of interest.

• …

CONFIDENTIAL


Estimate from proxy information – Example use cases

• Machine learning can be used to estimate hard to measure parameters easier to measure parameters as an alternative to the conventional observers and hardware sensors. This is especially valuable for cases in which the environment doesn’t allow for direct measurement. These estimators, also known as software sensors have been successfully applied in many chemical process systems such as reactors, distillation columns, and heat exchanger due to their robustness, simple formulation, adaptation capabilities and minimum modelling requirements for the design.

• These systems can predict unmeasured states such as concentration, temperature, heat flux, molecular weight and impurities from context parameters. An overview can be found in the paper: “Artificial Intelligence techniques applied as estimator in chemical process systems – A literature survey Jarinah Mohd Ali”

CONFIDENTIAL


CONFIDENTIAL

Agency

CONFIDENTIAL


Agency

AI systems can have agency, meaning they can act as an agent and learn directly from their environment. These types of systems are very good at learning to play games because they are continuously improving whilst playing. However, they have also found their way in robotics and some end-to-end autonomous vehicle applications amongst others. It should be noted that these systems are not often encountered when reliability and safety are required. They are hard to test because they learn continuously and the design can’t be frozen.

Reinforcement learning is the most popular technique in this space.

CONFIDENTIAL


AI, automation and robotics

At this point in time, for most robotics and automation applications, these end-to-end reinforcement learning models are not used in critical systems. When Robotics utilize ML techniques these are most often ML perception systems combined with some planning or optimization methods. These systems can be thoroughly tested and released in a controlled way.

CONFIDENTIAL


CONFIDENTIAL

Model complex systems

CONFIDENTIAL


Model complex systems

Machine learning algorithms can learn complex relations between a large number of variables. This allows for the modelling and characterization of complex systems with many variables.

The model is learned on measurement data of the process.

CONFIDENTIAL


Model complex systems – use case examples

• Stirred Tank modeling with Reinforcement Learning- ML algorithms were used to model the dynamics based on measurement data.

• Chemical reaction modeling learned from specimen data.

• Modeling plant operation, learned from data.

CONFIDENTIAL


CONFIDENTIAL

Optimization, search, planning

CONFIDENTIAL


Search / planning / optimization

Search planning and optimization are all about finding solutions in a large solution space.

Examples: • Automatic scheduling and

planning• Stock optimization• Production parameter

optimization• Vehicle routing• Information retrieval…

CONFIDENTIAL


Search / planning / optimization – use case examples (1)

• Price Optimization - Today, an AI algorithm could tell you what the ideal discount rate should be for a proposal to ensure that you’re most likely to win the deal by looking at specific features of each past deal that was won or lost.

• Process optimization - i.e. Machine learning to optimize process of continuous flow chemistry.

• Process Optimization – optimization of manufacturing process parameters using deep neural networks as surrogate models

CONFIDENTIAL



Supply chain management can benefit greatly from AI techniques for optimization: Improving forecast accuracy, optimizing transportation performance, improving product tracking & traceability and analyzing product returns.

CONFIDENTIAL



Artificial Intelligence for Inventory Management - Amazon examples1. Demand Prediction for Inventory Management2. Reinforcement Learning systems for full-inventory management.3. Robot automationYou may be using SAP, Xero or any other myriad of software for your inventory management. These can be integrated with Ai.

CONFIDENTIAL


CONFIDENTIAL

Information retrieval

CONFIDENTIAL


Information retrieval

Information retrieval is found everywhere, in the search bar on your phone, email and your search engine. Information retrieval can be used to search through multimedia databases, documentation, scientific literature and other databases.

Recent advances in AI like neural network encodings allow for faster and more intelligent search that goes beyond text matching. These technologies make previously unsearchable formats, searchable: • Searching similar images based on query

images.• Searching similar molecules based on their

chemical structure.• Search videos • Searching audio recordings• …

CONFIDENTIAL


CONFIDENTIAL

What AI can’t do… ( very well )

CONFIDENTIAL



• Dealing with the long tail of distribution• Learning outside the data• Explaining itself• Deciding - what probability is acceptable?• Reasoning – induction vs deduction• Designing itself

CONFIDENTIAL



Dealing with the long tail of distribution & Learning outside the dataAlthough modern machine learning algorithms are surprisingly good at predicting outside of sample cases correctly, most techniques require a representative dataset during training of the full input space. This means that ML won’t be able to learn a lot about samples that are very far from anything ever seen before. ( although the anomaly detection systems have ways to deal with this (see = “ recognizing deviations from patterns ). Explaining itselfDifferent methods have different levels of interpretability for humans. However at this point high performance methods like neural networks can learn very complex relations and patterns but have no way of explaining or providing insights into its learned relations. Deciding - what probability is acceptable?Many machine learning based decision tools will provide some type of probability output. For example it can output the probability that a chemical process is overheating. In this case, humans still have to decide what the threshold for action is and what that action would be. However, this is not always the case, one can let a ML algorithm learn optimal actions and thresholds in some cases. ReasoningHumans are very good at linear reasoning and reasoning by analogy with very limited information. There is a lot of research on this topic but AI systems are not yet at that point. Designing itself At this point AI algorithms still require a creator or designer. The A.I. expert is tasked with defining a good size and architecture of the AI model so that it can learn or perform the task at hand. There are a lots of research and first applications being created that try to automate this process but at this point A.I. experts are still needed in most cases.

CONFIDENTIAL


CONFIDENTIAL

AI approach

CONFIDENTIAL


“Let’s collect as much data as possible and apply A.I. later”

“We have a bunch of data laying around … “

“A.I. as a solution to everything…”

What is the right approach to AI in your organization ?

CONFIDENTIAL


Assess your organization readiness for AIStrategy

Readiness level 1 = initialReadiness level 2 = repeatableReadiness level 3 = definedReadiness level 4 = managedReadiness level 5 = optimizing

Adapted from the

AI awareness (in organization)

Legal Data

People

12

34

5

AI readiness requires a company wide commitment. Fill in this canvas to asses your readiness.

A legend of the axis is provided on the next slide.

CONFIDENTIAL


Level 1 Level 2 Level 3 Level 4 Level 5Initial Repeatable Defined Managed Optimizing

Strategy No corporate initiatives. Isolated.

Integration & cooperation in multiple business units.

Penetration of AI in all business units.

Evidence based process metrics regarding AI usage.

Continuous improvement & AI as a well-known and common strategy.

Data Scattered & unmapped data sources & tools.

Some centralized sources, tools or data lakes available.

Centralized data warehouses with mapped data quality & potential. Corporate standard tools.

Data management & value potential is known & reported on consistently. Enterprise AI architecture defined.

Active steps are taken to optimize monetization.

People Training & people is ad hoc & individual.

People development & regular courses.

Data competency & development frameworks.

Organizational structure, culture of innovation. Collaboration according to competencies.

Data literacy is a cornerstone of talent management with mandatory & continuous development for all relevant employees.

AI awareness(in organization)

Product owner Marketing R&D Higher management

Legal Scattered or unclear responsibility.

Awareness training. Defined & communicated responsibilities.

Clear responsibility with centralized oversight, enforcement & training.

Internal audits, mandatory reporting & penalization.

Legal compliance as an asset & unique selling point.

Assess your organization readiness for AI

Adapted from the Faktion framework

CONFIDENTIAL


Integrated approach to AI

AI algorithms don’t live in a vacuum. They interact through IT structure, with the physical world and humans. A good AI solution is a global optimum at all these levels to achieve the desired value proposition. Often the AI model is expected to solve the problem downstream of the sensors, IT and user components. This can lead to suboptimal solutions.

CONFIDENTIAL


When developing AI applications there are several levels of customization one can take. As a rule of thumb it is best to start with existing services ( top of the diagram ) and customize only when needed. Lower levels in the pyramid require more specialized personnel but allow more freedom to build custom applications.

Tools and frameworks

Integrated software components

SAP Leonardo

Salesforce Einstein

…

AI services IBM Watson …

Google Cloud services …

Amazon AI services: Amazon forecast

Amazon lex (chatbots)

Amazon recognition (computer vision)

Amazon translate

….

Custom development tools

Tensorflow (neural networks)

Pytorch (neural networks)

Google OR (search and optimization)

…

CONFIDENTIAL



SAP – Leonardo• SAP Conversational AI• SAP Data Intelligence• SAP Cash Application• AP Service Ticket Intelligence• SAP Customer Retention• SAP Predictive Analytics• …

CONFIDENTIAL



Salesforce Einstein

CONFIDENTIAL


Ai canvas

The AI canvas allow to analyze your AI opportunity. The goal is to bring all stakeholders together and fill in the 4 main blocks as best as possible: 1. The goal of the system2. The predict step. 3. The learn step. 4. The evaluation step.

CONFIDENTIAL


Predict Goal EvaluateLearn

Impact on decisionsHow are predictions used to make decisions that provide the proposed value to the end-user?

AI canvas

Machine Learning / inference tasksInput, output to predict & type of problem.

Making predictionsWhen to we make predictions on new inputs?

Offline evaluationMethods & metrics to evaluate the system before deployment.

Value propositionsWhat are we trying to do for the end-user(s) of the predictive system? What objectives are we serving?

Live evaluation & monitoringMethods & metrics to evaluate the system after deployment & to quantify value creation.

Data sourcesWhich raw data sources can we use (internal & external)?

FeaturesInput representations extracted from data sources..

Collecting dataHow do we get new data to learn from (inputs & outputs)?

Building modelsWhen do we create/update models with new training data?

Adapted from Louis Dorard’s Machine Learning Canvas v.04

Input

Desired output

Problem type

Prediction schedule

Prediction policy

Methods

Metrics

What

Why

Who

Methods

MetricsStatistical features

Expert features

Collection strategy

Model building strategy

Model building schedule

Workflow integration

Collection policy

CONFIDENTIAL


Predict Goal EvaluateLearn

Impact on decisionsHow are predictions used to make decisions that provide the proposed value to the end-user?

• Selection of molecules for further investigation. • Omission of molecules for further investigation.

AI canvas – QSAR* toxicity use-case

Machine Learning / inference tasksInput, output to predict & type of problem.

Making predictionsWhen to we make predictions on new inputs?

Offline evaluationMethods & metrics to evaluate the system before deployment.

Input Molecular descriptors Physico-chemical properties

Desired output Toxicity classification: Toxic/ non-toxic

Problem type Classification problem

Prediction schedule User (lab professional) initiated

Prediction policy Check data qualityAssure authorized user?

Methods - Lab validation tests for out of sample predictions- Expert evaluation

Metrics - Statistical accuracy error.- Receiver operator curve (ROC)

Value propositionsWhat are we trying to do for the end-user(s) of the predictive system? What objectives are we serving?

WhatClassify the toxic response of chemical agents based on a formal description of the molecule in silico. (QSAR for toxicity )

The in silico model is learned on a database of molecule descriptors and their toxicity.Why • Increases efficiency of

toxicity screening. ( cheaper and faster )

• In silico testing is safer than lab tests.

• Reduce suffering for lab animals.

Who• R&D team.• Lab personnel

Live evaluation & monitoringMethods & metrics to evaluate the system after deployment & to quantify value creation.

MethodsRandom sample tests on predictions verified by lab tests

Metrics% correctly identified:Confusion matrix

Data sourcesWhich raw data sources can we use (internal & external)?

Existing databases: PubChem, ChEMBL, …

FeaturesInput representations extracted from data sources..

Collecting dataHow do we get new data to learn from (inputs & outputs)?

Model StrategyWhen do we create/update models with new training data?

• Experimental : molar refractivity, dipole moment, …• Theoretical molecular descriptors: 1D,2D,3D,4D

• Fingerprints• Graph invariants• WHIM• ….

Collection strategy Lab testing champagne for out of sample products

Collection policy Good representation of different toxicity types.Good distribution over chemicals ….

Model building strategyArteficial neural networks for classification

Model building scheduleFirst cycle: Establish feasibility. ~200 datapointsSecond cycle: Increase + diversify dataset. Retrain & refine model

Adapted from Louis Dorard’s Machine Learning Canvas v.04

Workflow integrationR&D staff assessment of:1 – Ease of use2 – Clarity of result

*Quantitative structure-activity relationship

CONFIDENTIAL


CONFIDENTIAL

AI opportunities in chemistry today

CONFIDENTIAL


Applications of AI in chemical industry

• Manufacturing: ML is primarily used for monitoring equipment and controlling applications in manufacturing. ML algorithms predict failures in equipment and also predict the necessary maintenance of equipment. This results in reduced down-time which in turn optimizes production.

• Drug Design: Traditionally, drug designing is a long drawn complex process. But with ML tools such as self-organizing maps, multilayer perceptron, bayesian neural networks, and counter-propagation neural networks drug design has become less challenging process.

• Compound classification: Chemists spend hundreds of hours in compound classification when it is done manually. Furthermore, it is prone to human error and thereby not cost-effective. ML applications are being used for classifying compounds.

• Toxicity prediction: ML methods such as support vector machines (SVM) and artificial neural network (ANN) are widely popular to conduct R&D activities such as determining in vivo toxicity. Based on available in vitro bioassay data, ML applications can be designed to predict the toxicity of chemicals.

ML applications are also used by chemical companies in couple other areas supportive to manufacturing:• Demand prediction: For many chemical industries, the demand for products fluctuates throughout the year. For

instance, the demand for oil keeps changing every month. To further complicate, demand planning processes can be inaccurate and hence, too expensive for a company. But demand planning is a critical process that is required in manufacturing. ML algorithms, developed by data science experts, are accurate and are well-suited for conducting demand prediction.

• Workforce management: In chemical industry, skilled workforce is in great demand. Hence, hiring and retaining skilled employees is a challenge. Many companies focus on training managers, and employees to keep employee churn at a minimal level. ML applications are useful to predict workloads, identify departments with greater churn, predict employees who may leave.

• Business decision support• Inventory optimization

CONFIDENTIAL


Application of AI in Chemical Engineering & manufacturing

• AI in chemical process modelling.• AI in optimization of chemical processes.• AI chemical process control.• AI chemical process monitoring.• AI techniques in fault detection and diagnosis

of chemical engineering.• Predictive maintenance and fleet based

management for machinery.

• Quality control through automated systems ( computer vision )

• Automated Plant monitoring

• Waste minimization

• Manufacturing planning and configuration.

• Automation

CONFIDENTIAL


Application of AI in Chemical R&D

• Medicinal Chemistry and Pharmaceutical Research• Drug/chemical Design

• Target validation, small-molecule design and optimization, predictive biomarkers, identification of prognostic biomarkers and analysis of digital pathology data in clinical trials.

• Virtual screening (VS) for rational drug/chemical development. • Prediction in biological affinity, pharmacokinetic and toxicological studies, as well as quantitative

structure-activity relationship (QSAR) models.

• Theoretical and Computational Chemistry• i.e. Prediction of ionization potential, lipophilicity of chemicals,

chemical/physical/mechanical properties of polymer employing topological indices and relative permittivity and oxygen diffusion of ceramic materials.

• Analytical Chemistry • i.e. Neural network techniques with the aim to obtain multivariate calibration and

analysis of spectroscopy data, HPLC retention behavior and reaction kinetics. • Biochemistry

• Neural networks are being employed in biochemistry and correlated research fields such as protein, DNA/RNA and molecular biology sciences.

• I.e. e reverse-phase liquid chromatography retention time of peptides enzymatically digested from proteomes, prediction the stability of human lysozyme.

CONFIDENTIAL


Resource management

• Inventory optimization/management• Optimize throughput, energy consumption and profit. • Supply chain management• AI for increasing productivity and reducing costs• Advanced analytics planning systems

• Order/demand forecasting• (Human/material) resource forecasting

CONFIDENTIAL


Application of AI in business

“”The five largest Customer Relationship Management (CRM) vendors by market share in 2015 were Salesforce, Oracle, SAP, Adobe Systems, and Microsoft. These five companies make up almost half of the entire CRM market. All of them have been investing in their internal development of machine learning and AI, while also buying AI startups.” AI capabilities are currently available for each of the five CRM giants• AI-powered personalized marketing/experience – personalizing the

content each customer receives.• Predictive recommendations – using a customer’s data to recommend

products they would be most interested it.• Optimizing the selling process for representatives – opportunity

analysis of clients to create guidance to help close deals. • Help direct sales offer by finding patterns in crm that have high chance

of yielding new business. • New product introduction forecasting.• Chatbots for better customer service

CONFIDENTIAL


CONFIDENTIAL

Patent landscapeAI, ML …

CONFIDENTIAL


CONFIDENTIAL

The “AI” related patents are seldom application/industry specific. There is no significant patenting activity in industry 4.0 for the chemical industry specifics. For the fields of machine learning and data management processes, only 87 and 73 patents, respectively, have been filed during the last five years within the chemical industry.

CONFIDENTIAL


KEY TAKEAWAYS FROM PATENT SEARCHES

• Machine Learning related patents cannot be spotted in relation to the

chemical industry.

• Software patenting is by tradition very limited. By looking at patents related

to machine learning and data management systems in the chemical

industry for the last five years, we found 87 and 73 patents respectively.

• The key technological development values are related to threshold,

accuracy, efficiency, robustness, overfitting, noise. These concepts sit

around quantifying, measuring, delivering, processing and demonstrate the

willingness of the players to continuously improve and automate their

processes.

• The leading category into which machine learning and data management

are used for innovation within the chemical sector is the medical field, where

it is used for gene editing, antimicrobial agent preparation, data processing

of samples or automatic diagnostic process.

• On a global level, US is the main player 1/3 of the activity. The leading

European player is Germany with 1/8 of the global activity.

2614 patentsTimeframe: 2013 – 2018

Total factory control & predictive maintenance[Cross-industry level]

POOL 1

We have analyzed 3 different patent pools (see the results in the following pages)

87 patentsTimeframe: 2013 – 2018

POOL 273 patents

Timeframe: 2013 – 2018

POOL 3

Machine learningChemical industry

Programming tools or database systems

Chemical industry

CONFIDENTIAL


The openness pledge in AI (1)

In the A.I. researcher community an open source mentality has traditionally been applied. Even big tech companies like Apple, Facebook, Amazon, and Microsoft have all, like Google, released software their own engineers use for machine learning as open source. They have pledged commitment to openness in artificial intelligence. The big tech companies seem to have it both ways though:

“At the same time, these proponents of AI openness are also working to claim ownership of AI techniques and applications. Patent claims related to AI, and in particular machine learning, have accelerated sharply in recent years. So far, tech companies haven’t converted those patents into lawsuits and legal threats to thwart rivals. Google itself exemplifies the trend. In 2010, only one Google filing mentioned machine learning or neural networks in its abstract or title, according to a search of the USPTO database. In 2016, there were 99 such filings from Google and other Alphabet companies. Facebook filed for 55 patents related to machine learning or neural networks in 2016, up from zero in 2010. IBM, which has been granted more US patents than any other company for the past 25 years running, boasts that in 2017 it won 1,400 AI-related patents, more than ever before ”

U.S. patent filings in machine learning

https://www.wired.com/2015/01/facebook-open-sources-trove-ai-tools/

https://www.wired.com/2016/05/amazons-giving-away-ai-behind-product-recommendations/

https://www.wired.com/2016/01/microsoft-tries-to-one-up-google-in-the-open-source-ai-race/

CONFIDENTIAL


The openness pledge in AI (2)

The AI community is has a strong “ maker mentality “ of sharing code via platforms like github. Most big advances in machine learning algorithms and techniques are shared by universities and big tech companies. There are several reasons why this doesn’t necessarily amount to bad business rational:

• Tech companies often compete for market share, the tools they provide are an opportunity to bring makers to their environments, use their infrastructure (cloud platform) and to gather information on what is happening in the field.

• The competitive edge is often not in the machine learning algorithm but in the data used to train a specific instance of that algorithm. The data and trained models are not always shared, the techniques for training are.

Among the technologies that major tech companies have opened recently are:• Amazon’s Alexa, the voice-command response system inhabiting the company’s Echo device,

opened in June 2015;• Google’s TensorFlow, the heart of its image search technology, open-sourced in November 2015;• The custom hardware designs that run Facebook’s M personal assistant, open-sourced

in December 2015; and• Microsoft’s answer to these machine-learning systems, the prosaically named Computation

Network Tool Kit, made public last month, the latest addition to the public’s library of options for AI systems.

Initiatives like Elon Musk’s OpenAI are aimed at facilitating and boosting the open culture in AI.

http://www.theverge.com/2015/6/25/8845103/amazon-echo-alexa-voice-control-third-party-hardware

http://www.theverge.com/2015/6/25/8845103/amazon-echo-alexa-voice-control-third-party-hardware

https://www.tensorflow.org/

http://www.wired.com/2015/11/google-open-sources-its-artificial-intelligence-engine/

http://www.wired.com/2015/08/facebook-launches-m-new-kind-virtual-assistant/

http://www.wired.com/2015/12/facebook-open-source-ai-big-sur/

http://www.cntk.ai/

http://www.wired.com/2016/01/microsoft-tries-to-one-up-google-in-the-open-source-ai-race/

CONFIDENTIAL


One group, five brandsOur services are marketed through 5 brands each addressing specific missions in product development.

INTEGRATED PRODUCT DEVELOPMENT

ON-SITEPRODUCT DEVELOPMENT

DIGITALPRODUCT DEVELOPMENT

OPTICALPRODUCT DEVELOPMENT

http://www.lambda-x.com/

http://www.lambda-x.com/

https://www.verhaert.com/

https://www.verhaert.com/

https://www.moebiusdesign.be/

https://www.moebiusdesign.be/

http://www.load.digital/

http://www.load.digital/

Documents

Technology Watch AI and Chemical Industry€¦ · •Information retrieval •What AI can’t do… ( very well ) AI approach AI opportunities in chemistry today Patent landscape