Lecture 02 : Part II

[Music] welcome back this the second part of the lecture number two so we have been discussing about how you discover materials right and I Told You So Much thing about serend p as discovery of materials serend p means accident please remember Serendipity comes from very interesting aspects I can give one hour lecture on that let's not talk about it but Serendipity means accident something happens in accident which you don't want to happen nobody wants to have accident but it happens and then you discover something but you know in the world of AI rarely CPT will happen that's what we're going to see so let's talk about it what is the new paradigm Discovery Paradigm thanks to Ai and ml what can happen happen so first before 1600 why do you call 1600 because that is the time of Isaac Newton you know the Paradigm was empirical empirical means we know some connections relationship between variables and something but they are all empirical not based on scientific thoughts one example is you know rate of reaction people knew rate of reaction formulas much before but nobody knew how these formulas are are derived so they are like a empirical science then between 1600 that's why the Isaac Newton CES into picture 1600 and 1950 as slate as 1950 almost 350 years or do you have a second Paradigm that's why you have model and theoretical science a lot of development happened in thermodynamics kinetics many subjects especially in Material Science thermodynamics and kinetics were developed extensively in this 300 years time period or rather I'd say mostly 1700 to 1950 and that's why we call a second Paradigm this is a second Paradigm you can see here second Paradigm is listed here and there models are available we know the theoretical understand of subject but we didn't have computers the first computer actually was made in 1940s sometime I don't remember by IBM that was a you know very small effort it cannot do much calculations in those days but then came the explosion from 1950s to 2000 what is that that's the time we started understanding that any material performance of property depends on structure and structure depends on processing how do you make it so when you make a material you create certain kind of a micr structure or structure it can be connected to Atomic Structure also you can understand that using characterization and this structure will give you the proper properties remember we are talking about discovering material with Target properties so that's what is your a but remember property is not the extreme thing what is Extreme important thing is the performance of the device where this material is going to be used in a mobile phone you have a display and you know display is done by liquid crystal display or TFT whatever it is there are two many things available to in this two days LED is also available we have display by Q LEDs also so doesn't matter but you have to see it properly or your pictures videos everything should be seen properly that's the performance that's what you want performance is basically for a device and device is made of many materials each material has specific properties that's the difference between performance and properties so that's what came in picture in after 1950s so therefore 1950s and on we have a third Paradigm that's what people call as a simulation Paradigm or computational science Paran okay we have the density functional theories and molecular Dynamics theories available in the material science physics cond pH community and that's what given us completely new way of looking at things so if I go back starting from empirical to model based to computational aspects what do the change happen what do you think well only certain things change happen obviously between empirical and model based a lot of science has been added very clear because in empirical Paradigm we didn't know anything about what is happening scientifically we knew some relationships just like algebraic equations but we didn't know what is the origin of this algebraic equations remember if you have algebraic equation let's say have algebra equation if I have Al equation Y is equal to F of X1 X2 X3 dot dot dot something like this so that means what X1 X2 X3 are variables Y is a target so values of X1 X2 X3 will determine the values of Y so this empirical relationship tells did tell us that this y it dependent on these three variables so we did we knew that the factors on which Y is dependent on but we didn't know what is the exact relationship we have some empirical relationships that's the difference between first and second par second Paradigm actually was completely scientific B everything was SCI that's why even in 1900 quantum mechanics came into picture we have a lot of new understanding of the subject of thermics came into picture we know thermodynamics from equilibrium to non thermodynamics all kinds of things you can talk about about it and many other aspects in Material Science happened physical mety also under a lot of changes dislocations were discovered in uh understanding the deformation V materials and then you have all kinds of quantum effects came into picture nanomaterials CES in picture in later part of 70 or 80s so therefore we have a huge change from first to second Paradigm but can you use this scientific concept now thanks to development in computers computers means speed and the power of computers can you develop new paradigm that's what we are talking about a computational science or computational simulation time of Paradigm in which we can actually simulate structures so we can simulate even Atomic level configuration electronic energy levels we can also simulate and develop from very basic understanding what is happening in the atomic scale can it parate to the higher Dimensions from atoms to molecules to structure or micro structure grains can it calculate that that came into picture and that's why it's called a computional paradigm but you remember there is another Paradigm which I am missing here what is that let me talk about it a little later but you know with this competion Paradigm if I ask to accelerate the material science Discovery how will you do it tell me you have any idea well this is done by using obviously I tools or I different tools you know this picture tells you that six levels level zero to level five what are these what is level zero level Z is completely human driven correct so what is level zero level Zer is a typical human Centric experimental process and you know it's what happens till today in many Laboratories so students doing some experiments getting some results I'll say research actually resarch Scholars doing some experiment getting some results a majority of the experimentation is carried out by human so some pieces of equipments can be used you can use computers also provide signals to digital formats but there is essentially no exchange of information other than what is measured and processed by in human no exchange of information between machine and human equipments are used to generate data then data are used by human to create some kind of understanding that's the level zero that's the best level you can see as human Centric we do documentation then human look at the data analy the data to find out some correlation then little bit above is level one okay that's the level one you can see here so level one basically tells you what assistant level it is still human Centric or are human driven but some amount of AI came into picture what is it this sees the implementation of some type of assistance some type of assistance as robot black and Isaac tambin has written in this book accelerated material Discovery some kind of a assistance to the human task whether physical or digital doesn't matter it can be both assistant is minimal and it's confined only to S kind of single task not multitasking abilities as the case of ping machine or some form of automated data collection like that autom data collection can be done by a you know CNC machine or many other machine also at this stage one can expect only human intervention because you're collecting data you're not doing much so that's what it is it's accelerating factor for experiment is to increase what it will increase but it will increase very less then come the level two which is called partial automation what is it at this stage experiments workflow is becoming much more automated that's what it is so you have automated many many things you can actually do an experiment in absence of you you can set it up an experiment will be done and data will be collected and with various experimental procedures capable operating over extended period of time automated ones this is called consider as a high throughput experimentations that's what we call it high through throughput experimentation in this today's world High throughput means you generate a lot of data in small high throughput experimentation it generate a lot of data in small time if human carries out this experiments it will take a long time that's why we're talking about tile and error here it is not the case here we have automated some part of it so the data generation is faster and experiments are very quick so that's called high throughput experimentation stage and obviously workflow Pace will increase quite a lot while the workf flow is in in quite a lot of increased incorporation of physical or digital means are very important the human is still in control of the whole thing human has not gone you can see here human intervention is still there so we have to go a little bit further that is called level three what level three level three is basically conditional automation very interesting name conditional automation automation means you automated it but it's a conditional things so what is it actually this St this stage sees application of AIML so in stage three only you are started using a IML that's what you are analyzing your data using all kinds of machine learning algorithms or models and you know we can create optimization Loops which are coupled to the platform automation so we can use statistical tools and deep learning algorithms if you want into the workflow a system will be capable to make its own decision so 0 to2 system was not making own decision human are making decisions but here system will be capable to do some exp some kind of a decision making process and you know based on the whatever data available this can be considered actually B best level for self- learning so we can write down at this as a best level for self- learning best level for self learning machines will start learning at this level level three and that's why this is we call starting point of using AI so it took three levels to achieve that that's what it is now but still knowledge is still a key to successful it drive the experiments effectively so therefore human is required it cannot be completely decoupled but human intervention is less because machine is doing some calculation or modeling to get some kind of a sophistication that's what it is so please remember this is called conditional automation it is automation is there but it is automated but is conditionally automated now come the level four thing which is very interesting it's called partial autonomy why it is called a paral autonomy let's see well at this stage system starts to truly become autonomous as it begins to inform properties and relationship from outside sources like literature past experiments it is trying to build that okay so it begins to infer properties of the system and then find out relationship from outside so literature past experiments why to truly plan and execute autonomous experimentations if you want to do experiment in space that's what you should do human will not be there so there paral autonom is very useful coupling Theory empirical data will allow the system to understand INF for chemic processes using knowledge to more effectly search the material landscape human role is very limited that's what you see human role is very limited now comes the last stage a last level that is called the full autonomy or level five what is it this includes a truly autonomous researcher where human can effec will be removed from the system some people say this is the bank if that happens AI will take over and then human like us will have no role in this world and that's not it happened that thanks to the literature human can effectively remove from this picture platform is able to complete all task of Discovery or reporting and it determines what experiments what simulation what theoretical approaches need to be applied to optimize even active learning process so it can use its own thought process to do that that's what it is and you know furthermore the stage is also represent truly autonomous researcher so that's called true autonomy so true autonomy and it can actually do all kinds of things so we want to achieve that final goal for human to achieve that it's completely I even even the properties interference full system control everything is a they'll do it but you know this requires this is the required paradigm shift so that's what I bought this before I telling you the last Paradigm fourth Paradigm I want to show you this so therefore from Level zero to level four we already have discussed a lot of things some part of things we have already discussed previously level G is basically trial and error level one some kind of a you know these things which are like a using intuitional things level two maybe much better intuitions but Serendipity will somewhere come into picture Serendipity is only for accidents people who are who can accommodate accidents in their life for them it is important so therefore what it follows is that role of laboratory automation machine learning data infrastructure will play a very important role to achieve this true autonomy autonomic pictures in this world and that is something which we'll learn in this course I hope you can read more about this from this uh chapter in the book called accelerated material Discovery and this chapter was written by Robert black and Isaac Tamlin and please read properly before we talk even more about this last Paradigm which is basically hint to you you know typically how do you traditionally screen materials I given you some example suppose you are using you want to create a you want to search a material for wire or basically transmission line Hy tension transmission line or you want to Choice hit sync typically if you don't have anything with you how to search first you select few new candidates and then you synthesize that analyze that test it and then screen for adverse skin SP which are not working and then take those which are another possible set again do the experiment that's what is called traditional material Discovery Paradigm so this requires you to do so many things time consuming aspects which may be costly also so instead of that can you do virtual skinning well that's what it is so I have given you this traditional screening already and then I'm giving you new program called computation Material Science so what you do you select a new candidate you screen in the first case that's what you do we add a screening parameter here by Machine learning or by automation AI tool that's what you do and this will go on this Loop will go on and then it will give you some choices which are likely to be successful based on those some choices use your intuition and synthesize those materials and after that analyze and then test for Effectiveness you you're bound to get a success I can tell you so that's what it is so therefore what is happening here is this in this chart we can show it initially you have lot of choice so by using Automotion or AIML approach or compression material s approach whatever it is you can talk about it you can reduce the number of choices so that you need to only prepare few materials synthesize few materials so time is saved energy is saved saved cost is sa and from this you analyze and then find the effectiveness and find out the right one that's what it is okay so this kind of things are very common today so therefore it is from addisonian to th throughput to Target design approach that's what I talked about in the last slide this is called Target design approach okay Target design that's what we talk so in case of addition approach you are going to only do hit and trial you have no idea what you're doing you try it find out if it works it's fine if it doesn't work you change it high through to experimentation what did I told you you can actually you remember let me go back this is what the high throughput experimentation is in the level two the partial automation you can fit as many gadets as possible to your experimental setup and get data faster experimentation a lot of data availability that's what is this fast experimentation that will make you things faster but will it allow you to go to Target design approach no answer is no in order to achieve the target design you need to do something else that means you want to achieve a targeted property for a particular material synthesis but you know on the top of which you have another thing to be added that is the manufacturing attributes I can make a material small quantity in lab but when you want to make large piece or in large numbers in order to manufacture and sell I have something else to be taken care of so I have to talk about product attributes like density hardness Thro toughness many other things and as well as manufacturing abil like viscosity glass transition temperature liquidize temperature cost CO2 emissions many things environmental impact all these must be taken care of so these attributes are never considered when you're designing a material a discovering material in the lab so that's why if you want to have you know glass you want to wear glass glass M you want to design new glass compositions you see there are about 10 52 possible glass composition available and only 400,000 glass compositions I explored over the last 6,000 years so that's 10 52 is similar to number of stars in the our Milky Way galaxy and we cannot explore it is impossible that's a massive number so last 6,000 years you only export 400,000 so that is 4 power five 10 52 is much much higher so therefore you have to use what is a principle called optimization principle and that's why it is required so I'll talk about more of this later but let me talk about this slide which the last slide for today's lecture so we have talked about empirical science in the first Paradigm second parad was model based theoretical app science third Paradigm is a computational material science computational science or simulations fourth one is basically the data Dem science or is basically Big Data Paradigm you must have been hearing these aspects quite a lot so whenever we purchase something online store that is stored in the database so next time if you browse the same uh you know online store they know what are your choices that's what is the use of big data is so Big Data uses all kinds of machine learning models to predict from 2000 onwards from 1950 2000 it was computational Material Science 2000 onwards we have what is known as a big data div approach or fourth Paradigm I don't know whether we have some more exist or not that fifth Paradigm was there or not that will time will tell but at least we are into the fourth Paradigm of our scientific new material Discovery so up knowing all these aspects you must be understanding that how do I know the fourth Paradigm without knowing the third Paradigm so well many of you may not have a complete picture of computational science or simulations many of you may be experimentalist don't worry you don't need to know about it computational science is more like use of these equations in a computer to solve something but data Paradigm is different use data to predict something so they are connected but they are not so much dependent on each other that you cannot learn fourth Paradigm without knowing much about third Paradigm that's my advice to you so please do not worry about it many of things will be connected to each other I'll make sure it happens so that you learn all these aspects so with this let us stop uh the second lecture I hope you have understood many of these aspects in the next lecture I'll also do some recap so that this Concepts will be much clearer to you thank you [Music] [Music] [Music]