Lecture 40 : Part 2

Students welcome back to the second part of lecture number forty and this is what is going to be your last part of your lecture and as I put it on the slide how do I sum it up. Summing up the whole course is a very difficult job. So I was thinking about it how to do it as you know summing up means how to put things together so that you understand everything. So next 20 minutes or so we are going to do that so that you can keep it as one last video how what are things we learn. So first we started about definition of AI, how AI is important and we don't need to explain it to anyone right even the kids also know why how AI is important. But they don't know that one thing that AI has limited capabilities that is something which you have to realize AI has two parts one is called machine learning and is called deep learning. Machine learning gives you structured data, deep learning handles unstructured data like text images right and you know machine learning in which machine can automatically learn the patterns from the existing data and we can actually do a lot of things using AI we can do analytical problem solution, learning, observations all kinds of things. And there are two types of AI one is called narrow AI which is that is what we are doing it we can detect faces, plugging online contents, customer queries, online chatbots but generally AI is still a concept today we are coming to reach that I don't think it is easy possible in a very small time frame it is envisioned to be adaptable and reasonable as human being remember the words adaptable and reasonable. But you know why you want to use AI basic reason is to discover and design new materials and what are the traditional way of designing new materials first is trial and error which we have been doing for ages right we try do something fail learn from the failures and then go back again change the conditions and this approach is a time consuming sometimes it will be like taking edges not possible for life of a human being to discover and this is what is called Edisonian approach because Edison and Tim have tested more than 3000 designs of the bulbs between 1978 and 1878 and 1880s within two years they tested more than 3000 designs of the bulbs so you know imagine how many people are involved and how many times they failed to succeed you need to fail that is what is the real problem by Edison approach but you know that is no longer possible so you have to think about something new then you have to understand how this subject or knowledge has changed from first paradigm to the fourth paradigm first paradigm was empirical science everything was empirical that is below 1600 before the Galileo's time and then come second paradigm which was basically model based science correct and finally after 1950s what we have basically something where you can correlate Arsena subject correlates taxa properties with performances and this third paradigm came in 1950s what is known as the computational science which are basically simulation based and there you can actually do lot of simulation using computers basic functional theories molecular dynamics and this can provide us new ways to design materials but that is not all then what you have is basically machine I know AI based techniques what is called a fourth paradigm this is what is called a fourth paradigm you can go back to my lectures you can find out that there you can handling data science we have lot of data how to analyze the data try to find out the trains just now we talked about in the last part or first part of this lecture so that is what is called big data science and this is something which we are all living in we can have predictive capabilities clustering relationship mining anomaly detection everything is built in this now you understand the meaning of the sentences after the all the lectures are delivered you can understand that so this big data driven science is something which all of us going through how to do that so this paradigm shift from first to second to third to fourth especially the third to fourth is very rapid in last 50 years between 2000 the third paradigm has started now we have gone into fourth paradigm so you know this is something which you wanted to learn that is what I started with but you know we are doing materials if you want to do materials modeling you should know what are the basic thing of material science basic thing of material science is the correlation between processing structure and property and we spend quite a lot of lectures almost like a 12 to 13 or 14 lectures on this aspect try to understand you what are the basic aspects of material science so that you can design new materials with target set of properties discover new materials and there we talked about fundamentals of structure properties and all these stuffs crystal structures atomic structures microstructures nano structures atomic structures and we talked about different properties mechanical physical magnetic electronic electrical and we expected you to learn a full understanding of this correlation between structure properties that is not easy remember that is that is the basic of material science we have to integrate this basic things with the big data science and that is my main motto when I design this course so that you can get lot of these basic concepts clear and apply this or take the clue from this basic concept apply this in the data science and you know materials can be anything but are they useful that is what we talk about engineering materials engineering materials are once which human extract or make for any utility so you can have different kinds of material like metallic materials ceramics polymers composites they are all called engineering materials where they are useful then you know do that you need to basically understand basic of material science and also you need to also understand the correlation between the structure property and processing so that you can apply this in designer materials and basic thing is the microstructure what is that is what is controlling the material properties and there is a key link between this synthesis or processing of materials and the properties and structure means starting from atoms crystal structures grain structures or finally microstructures this is what is your structures you have grains boundaries defects everything is going to come into picture you are going to have single phase microstructure multiple microstructure like steels you have alpha iron plus cementite in pure iron you have only grains of alpha iron so this is the difference between that when you have two phase or three phase microstructures distribution of the phases are also going to play important role interface between the phases also going to play an important role then you have defects you cannot have perfect things in the world nanoparts are perfect so imperfection the crystal structures are very very you know important you can have point defects you are going to have line defects like dislocations you can have surface defects or 2D defects like twinning all these defects play very important role in materials and we talked about lot of engineering materials metal ceramics polymers you can see non metals composite ceramic plastics many things but you know metals are the largest group you can have steels you can have aluminum alloys you can have titanium alloys I try to avoid it because that is a massive thing and then you can have ceramics ceramics are compounds the metal and non metals they often form oxides carbides nitrides they have much different properties than the metal materials they can have very high melting temperatures very high compressive strength but they are very little they are low thermal electrical conductivities right tungsten carbide and zirconium carbide aluminum oxide zirconium nitride all these examples are given these are there you can have hydroxyapatite, kaolinite, sialene then you have whole lot of things in polymers you have thermoplastic, thermosets, elastomers you can also have amorphous materials presents like glasses many of the polymers actually amorphous so the amorphous materials are widely present in the world and we cannot avoid it compared to the metallic and ceramics amorphous materials are mostly either glass or polymers so amorphous materials have random structures okay glass is a amorphous structure of sodium oxide, calcium oxide, silicon dioxide right remember that and if there is no grain boundaries that is not order assuming atomic elements in the glasses that is what is going to happen so in a in a in a polymer you can have completely random chains of molecules you can have crystalline plus amorphous structures also but that was the different structures you can also have different property like tensile compressive CR in mechanical properties you can have electrical properties like conductors, insulators, semiconductors we discuss different models of that and you can have you know metal ceramics, polymers, semiconductors again then you have magnets magnetic properties are very important right they actually provide you a lot of interesting insights of that and then we went into what is known as a processing conditions operations different kind of processing operations because those are the structure properties then you have talked about processing conditions shaping operations how do you shape a component how can you improve properties like heat treatments or something you can do surface treatment also like clean, coat, treat, deposit materials and shaping operations are four types we talked about a solidification processing, particular processing like a sintering, deformation processing and material removal process like machining or something right. So in solidification processing we have to pour a liquid metal into a cavity and solidify you can see that in powder processing we take a powder as a source of material sinter them to form a component and in deformation processing we take any bulk material deform it plastically to give a shape machining is the removal process machining means we cut something remove materials to give a shape right so these are the different shaping process we discussed and you know more important thing is the bulk deformation processing like rolling, forging, extrusions, stress forming, bending, sheet metal forming all kinds of things possible we discussed quite a bit of this powder metallurgy we take the powders we process characterize the powders then we do pressing and sintering and we can actually do other kinds of techniques also. So by pressing we create a green compact and this green compacts are then basically sintered to obtain a this sinter products. So this was about the basics of material science which I covered in last five minutes okay but it will take you one full course or maybe more than that so please look into different code structures and NPTEL or other fora where you can actually get a knowledge of that. Then comes the concept of machine learning how it is different from your conventional approach like trial and error methods and you know we can you know that we have trial and error best discovery of design materials we can also have third paradigm where you can use simulations like quantum mechanics, molecular mechanics all kinds of things you can do that. But we can also do high throughput skinning we can skin the materials and get lot of understanding of that right and machine learning models actually methods can be utilized to speed the material development cycles and machine learning is something which we can learn from data I discussed a lot about it how the task is important there are two types of machine learning supervised and unsupervised. Supervised means it develops predictive model based on the level data which has both input and output unsupervised means you do not have level data you have only either input no output or output with no input side. So you can actually use classification regression based models for supervised learning and clustering and for unsupervised learning this is the machine learning workflow let us now discuss about this supervised learning actually requires training and they are by level data level data I explain you how level data can be obtained from supervised learning. So first model we talked about is simple regression the statistical model the utilize utilizes one quantitative independent variable x to predict quantitative independent variable y you can have many quantitative independent variables also and we can also do the analysis easy ways to do is a linear model we can have lot of nonlinear model possible linear models will have some random error also you can see that and this actually linear combinations of these variables we can do classifications I talked about lot about classification problems machine learning can do a classifications easily and you know classification based on discrete values levels or you know even you can also do other kinds of levels also possible and you can do numerical predictions also like credit known medical diagnosis for detections all kinds of things. We talked about different kinds of classification problems like decentries, support vector machines and random forest and also Kenya's neighbors right decentry is something which we started discussing about it this example like whether can you play tennis depending whether right sunny or work as rain and we talked about SPM support vector machine how can you have a maximum margin remember I talked about roads right maximum margin is a linear classifier and here you can actually separate data with the maximum margin how we can do that is what is called linear SPM does and this is something which widely used in material science you will be using that. Then we talked about quite a bit of the artificial machine learning because this is important for this it is basically uses the concept of how the brain works how the brain has humongous networks how it works that way and it is basically based on dendrites, cell bodies, synaptic and axons and that is a typical elements of any axon we can actually have input variables we can put weights we can then add it up and then we can get we can put a transfer functions activation functions on these additions and get output. So this transfer and activation functions can be different types which you have discussed already in your lecture you can see that how this is this can be done your inputs and then you can weights multiply the inputs to the weights add it up and put activation function and get output that is what is what is called the ANN does. You can also have k nearest neighbors models this is you know it is basically we can do a majority category levels and very easily you know then we talked about clustering which is nothing but unsupervised models and you can have centroid models connectivity models tensoid models it is a task of dividing data into meaningful groups that is what is called clustering and it is aim to forming groups of homogeneous data points of a heterogeneous data sets. So centroid model is widely used k means algorithms you can have hierarchical algorithms other things also possible let us not talk about here quite a bit because this things can be easily obtained and you can also have density based connectivity based distribution based clustering also possible clustering is very important please do spend time on this learning the clustering algorithms. Then after machine learning we into deep learning deep learning is typically based on artificial neural networks you know and that is what we spend lot of time on how to do that you can use computer visions it can use NLP's to help your dynamics that diagnostics and automatic diving also possible using deep learning models, but it requires lot of data without that you cannot do you can have computer visions you can have also NLP as well as reinforce learning you can have you say agent to learn in for you that is what is called reinforce learning and there you have when you have talk a lot of data you can talk about different types of data like sequential data 2D data tabular data right and these each of these data are very important when you model and for deep learning you will use lot of data small data machine learning is good deep learning is not good you need a certain amount of data to a good DL modeling. So, we talked about ANN, CNN, RNN, autoencoders, DVN and many others advanced models also you know that and you know these models actually depends on feature engineering hardware dependence data types and they sometimes are called back box type they very difficult to interpret this models. So, this is the chart we talked about it we talked about supervising models like MLP multilayer perception, CNN, RNN, we talked about unsupervised like JAN and AE autoencoder, SOM, RBM, DVN and we also talked about hybrid models mixture of these different kinds of models right we spend lot of time in understanding deep learning models and you know I do not want to discuss each of these CNN is something which is widely used for you know image processing models that is what we did. Then we talk little bit about advanced models like advanced DLs which are widely used one such model is NLP natural language and from there we went to material informatics right that is what we talked about in last four five lecture NLP is something which is going to come more in material informatics how to generate data and do that and NLP has different kinds of phases lexical analysis, syntactic analysis, semantic discourse integration and pragmatic analysis and this is a very powerful tool which will come back. And finally, we also talked about it why you want to do that we have materials design at different length scales you can have properties perform structures and processing and these are connected to get an idea about performance right that is the key idea. So, when you talk about integration of data science with these things what is called the length scale first you start with the Calford and Calford provides you calculation of phase diagrams that gives you all kinds of property like entropy, enthalpy, qcp energies, diffusion coefficients, heat capacities and also it can provide you sometimes calculation provides you many other data. But microstructure model is very important that is what you need to use phase beam modeling and phase beam modeling you say interface there is a diffuse interface model where you can actually add a interface term and do that. Then finally, you know one of my colleague from other IIT, IIT Kharagpur is talked about computational chemistry and material science like electronic structures, DFT calculations as well as your molecular dynamic simulations like this is what is your Newton loss of motions are used for molecular dynamic simulations. Finally also we talked about it some beta of data preprocessing and other quality data cleaning, data integration, reduction and transformations. And why you need that we need accurate data which is complete, consistent, timely, believable, interpretable that is why we need to do data processing. Please remember that data processing is the critical step when you want to do a machine learning based models. I do not want to go in details of that finally, you have to do a model score that is what F1 score does how a model is performing good or bad or not. And this is nothing but a harmonic mean of the precision and the recall and we talked about it how what is it. And at the end I talked about a material informatics which is you know based on information data science and it is very important tool now a days to discuss about it. And as I told you we are going to develop a new course of material informatics may be after one year or so you will get to know about it. And this course we will talk about details of the material informatics because data science is going to be very important in material science also. And I will discuss all these aspects I do not want to go back into that. I just like to say you that I hope you have enjoyed this different lectures of this course and this has provide you some basic concepts of the machine learning and deep learning and informatics which on which you can built on. This is like a foundation and now on the foundation you can grow your knowledge and gain more and more understanding of that. And I hope also my TES has given lot of time to answer your questions, explain different kinds of you know knowledge these concepts by giving examples, by giving some kind of a programs or so we tried best to do that. And you know this require science this require lot of tutorial kinds of things, but it is not easy to do tutorials in a in a online based course, sometimes very difficult. Normally we do tutorial when you teach the course in the offline mode, there we spend some time on tutorials, explain different examples to bring different example to explain different concepts, but it is not possible here so likely. But I must tell you the participations of these kind of you know online descriptions were not very enthusiastic. And TES have also informed me that out of 8000 students hardly about 1000 students used to participate that is another important aspect we have to work on, how to make it more adapt more you know applicable to you. And last thing I want to tell you that I enjoyed developing this course, because I learned a lot and this is a very interesting concepts on which we will built on more and more courses. With this let me give you big thank and end this course by telling you that take care and do well. Thank you.