Skip to main content

Optimizing EV Charging Experience: Machine Learning for Accurate Charge Time Estimation

Summary

  • ML-based charge time prediction: Scalable machine learning models accurately estimate vehicle charging duration, accounting for dynamic factors like environment, battery state, and charging conditions.
  • Rule-based limitations overcome: Traditional rule-based methods fail with varying environmental conditions—ML models capture complex interactions for real-time, accurate predictions.
  • Vehicle control integration: Real-time charge predictions enhance both vehicle performance monitoring and user experience by providing accurate remaining-time estimates.

Optimizing EV Charging Experience: Machine Learning for Accurate Charge Time Estimation

Watch: Optimizing EV Charging Experience: Machine Learning for Accurate Charge Time Estimation
Talk by Mohammed Farag, Senior Manager, Machine Learning & AI, Rivian Automotive, LLC ; Sihang Chen, Sr. Machine Learning/AI Engineer, Rivian
Accurate charge time estimation is key to vehicle performance and user experience. We developed a scalable ML model that enhances real-time charge predictions in vehicle controls. Traditional rule-based methods struggle with dynamic factors like environment, vehicle state, and charging conditions. Our adaptive ML solution improves accuracy by 10%. We use Unity Catalog for data governance, Delta Tables for storage, and Liquid Clustering for data layout. Job schedulers manage data processing, while AutoML accelerates model selection. MLflow streamlines tracking, versioning, and deployment. A dedicated serving endpoint enables A/B testing and real-time insights. As our data ecosystem grew, scalability became critical. Our flexible ML framework was integrated into vehicle control systems within months. With live accuracy tracking and software-driven blending, we support 50,000+ weekly charge sessions, improving energy management and user experience.
Databricks Named a Leader in the 2025 Gartner® Magic Quadrant™ for Data Science and Machine Learning Platforms: https://www.databricks.com/blog/databricks-named-leader-2025-gartner-magic-quadrant-data-science-and-machine-learning
Build and deploy quality AI agent systems: https://www.databricks.com/product/artificial-intelligence
See all the product announcements from Data + AI Summit: https://www.databricks.com/events/dataaisummit-2025-announcements

Chapters

FAQs

Why are rule-based charge time estimates inaccurate?

Rule-based methods use fixed thresholds and don't account for dynamic factors like ambient temperature, battery chemistry variance, power source stability, and vehicle-specific characteristics.

What factors does the ML model consider for charge prediction?

The model incorporates environmental conditions (temperature, humidity), vehicle state (battery health, chemistry), charging infrastructure (power level, connection quality), and historical charging patterns.

What is real-time machine learning?

Real-time machine learning refers to ML models that generate predictions on incoming data with extremely low latency, typically measured in milliseconds, enabling immediate action based on model outputs. This video demonstrates real-time ML in the EV charging domain where charge time predictions must be generated instantly as charging sessions begin, accounting for current conditions rather than using pre-computed estimates. Real-time ML is essential for interactive experiences where user-facing applications need immediate model predictions without perceptible latency.

How do I automate model training, testing, approval, deployment, and retraining?

Automating model training, testing, approval, deployment, and retraining requires a platform that can orchestrate complex ML workflows while capturing data lineage and maintaining audit trails for compliance. Databricks enables this through Delta Live Tables for data pipeline orchestration, built-in ML workflows for model training and evaluation, and Unity Catalog for governance throughout the ML lifecycle. The platform can automatically trigger retraining when data quality metrics change, detect performance degradation, and roll back models without manual intervention.

Which platforms support reusable features for batch and real-time machine learning?

Databricks provides the best platform support for reusable features across batch and real-time scenarios by using Apache Iceberg as a unified table format that serves both offline feature computation and online inference pipelines. Rather than maintaining separate feature stores, teams can store features in Iceberg tables and access them efficiently for both batch training (historical data) and real-time inference (current conditions). This unified approach eliminates feature inconsistencies that plague systems using separate batch and real-time pipelines.

Full transcript

0:00 Introduction to Rivian

Hi everyone. So excited to be here. Uh I know it's 5:20. Um so a special thanks to everyone who have chosen to stay longer and not head straight to the happy hour or lunch or dinner. Sorry. Um this is a real dedication. So happy um for this. So I promise that to make this session worth your time. Um, today I'll walk you through how we optimize our EV charging experience at Rivian using AI and machine learning. So, Rivian's mission is to keep the world adventurous forever.
Quick show of hands, how many of you have driven or drive a Rivian? Awesome. Great. For those who haven't tried one yet, I hope the next video will give you a glimpse of what makes Rivian different. This platform approach means new features or improvements whether it's related to your charging time, longer range or even new powertrain drive settings can be rolled out across the entire fleet. Not just one model. At the core of scalability is our datadriven development.
Uh by using machine learning models powered by real world data augmented with over-the-air updates, we are able to ensure that every single vehicle on the road gets smarter over time. That's how we are able to move from gen one to Gen 2 to affordable mass market vehicles all while keeping software, AI, and data at the center of our journey. So let's start with why fast charging matters so much. First, so many for so many EV drivers, fast charging is the difference between local driving and road trips.

2:24 EV charging challenges

It enables longd distance trips and relieves the classic range anxiety. Second, knowing how long you will need to charge is very important and crucial because it allows you to avoid idle fees. Plan routes better and actually you use use their charging use your charging time well. But this is not an easy ride. There has been a lot of challenges especially with charging speed isn't just about the station. It's about the weather.
It's about the vehicle, the battery and its chemistry and even the battery configuration. This is all these challenges are exactly what we are trying to solve. In short, if you want EV adoption at a scale, you want to make sure that fast charging is predictable and as easy and accessible as possible. This is our motivation and our challenge. As we said, charging experience is the cornerstone for EV adoption. One key aspect of charging experience is the time a customer spend charging their vehicle, whether they spend it on the in the car or outside.
Imagine two customers arriving to the same station at the same time, both have identical needs of charging 10%. Even though their needs are identical, the time they will take to charge their batteries will be different. This is because the charging is depends on several factors. Some of them are short-term like state of charge and state of temperature of the battery and others are long-term like state of health of the battery.
Same time the weather plays an aspect environmental condition and the network itself plays an aspect there. This is a multi-dimensional optimization problem and it was a prime candidate for a machine learning based model to augment our existing invested model to provide a hybrid solution to enhance our customer driving charging experience. This is the endto-end charge time estimation architecture. We start with collecting data from the vehicle.
All the historical data helps us sorry all the historical data helps us to develop our features do model training selection optimize and deploy the model. This is a pass on the top. The other passing down there this is where our real-time data gets passed to a cloud service orchestrator where we have a streaming engine and aggregating uh aggregation engine. Both of them provide the model with the real uh accurate input so that the inference is being passed to the vehicle in this journey.
Those are all the tools that we have utilized from data bricks. This is our five stages for cyclic development in the ML life cycle. First stage we start with collecting data doing our exploratory data analysis and feature engineering. Moving from that stage to model training. This is when we use AutoML for Revit prototyping development of hyperparameter tuning and load into our experiment metrics. Continuous integration is one aspect here that triggers that step for us.
Stage three is model registration. This is when we select the best model, we version it and do batch testing as well. Stage four is model deployment. whether it's going to be cloud edge deployment also how it's going to be rolled out shadow mode AB testing canary testing all of those are part of our CD process that's also triggered as part of the overall system our last stage is monitoring and this is when we monitor drift and evaluate for retraining I'll pass it on to Canden to go in all the details for all of these stages thank you Okay, nice applause.
Thank you, Mo. So, I'll start by starting this journey with you guys. We go through each stage. That's my favorite yellow. So, the first one is data processing, right? Uh for data processing, we think for the whole model life cycle, it's always defining what's the business need and then you need to understand okay, what kind of data you want to take from the system and then ingest it into the model. And from this figure you can see like half of it is like stopping before the data right and I think 60 60 to 70% of the effort will be focused on getting a good data.

6:37 Data processing pipeline

So a good model you have to have good data or a good data you may also have a bad model but a bad data you definitely have a bad model. So in that case you need to spend a lot of time to make sure the data we input into the model is good is gold. So data is the backbone of this model and uh to understand how we retrieve those data like basically for all the problems you need to understand where the data is from when you receive the data how do you process the data then how do you save the data like where do you write the data in this case we utilize the delta live table it's good for batch and streaming and supporting like both strategies it has linear tracking basically you can know like for each of the data what is upstream table that's uh that they are relying on and it has efficient incremental processing.
Uh so every time new data comes in you're able to process it instead of doing all the historical data at once which is which is going to be very costly and time consuming. Then time travel is very important because if you did something wrong like in the new version you can always go back to the previous version and reset it. And finally data quality it has a lot of functions saying you want to check which column is always null which column is always zero is the distribution making sense is it skewed stuff like that.
So on the right we have these two figures uh talking about what is batch and streaming because that will be important concepts for our next slides. The first one is for batch processing. Uh what does it mean? It basically means we need to create this specific clear boundary of how do we want to slice the data right and we need to slice the data for us like on a daily basis like time series level or weekly basis is important.
We want to make sure like each day we get this new data comes in and we are able to process it and write the data into a table. Second is streaming which is important if you are if you care more about the event that happens lively. You want the result instantly. So for streaming is event based may become one event two events or more and then it will go into your processing logic process the data and finally finally write the data into a table.
So that's how we understand okay you know where you ingest the data you know how you process the data and you know where you want to save the data. So the next step is how you govern the data. So we utilize unity catalog which is uh provides us this centralized data governance and management. Uh then we have access control to see okay which layer of data we want to show to other teams which of layer data we want to just use internally for some development.
And finally you have this lineage tracking. So we we have this medality architecture. If some of you guys have attended the training in recent days, you may be familiar with this concept. But for us like each medallion architecture have different like specific use case. On the very top we see we have bronze layer which is we ingest the data from the vehicle and uh we develop this aggregation logic and then we write the data into a bronze layer.
So that's like raw data from the vehicle and we create aggregations like session definition then we write it and then it goes to the silver layer where we basically stick a lot of vehicle attribute data let's say the battery pack or what vehicle platform that is and it's managed by DBT it will be appended to the bronze layer data and finally we can apply our domain knowledge and apply some aggregation logic uh of the silver layer data managed by DBT and finally write the gold layer data into a So that's for uh this medallion architecture.
Right now we have uh this go layer data which can be used by downstream teams for analytics or like uh machine learning development. Then after you have you know you know where data is from you know how to write it you have this medallion architecture to give you to to basically send you to a stage where the data is good to use. You need to make sure you have the job scheduler where because for us we care about the new data comes in every day, right?
So we need to set our pipeline using job scheduler and automate this whole process. It's mainly applied to the bronze layer. So for job scheduler there are several functions. First one is parameter definition. You can define you know where we want to save the data and what is the environment. Is it testing environment? Is a prod environment? What kind of vehicle fleet we care about? Is engineering customer etc. And then we have task hierarchy and cluster configuration basically to make sure the cluster the configuration of the cluster is right and is assigned to the right team.
Then uh we have you know data every everything is ready you have this data coming in the database you want to make sure the downstream user is able to use it in a very efficient way right for me like I work on machine learning every time I I I do the model I need to make sure when I query the data is fast so we utilize uh liquid clustering so liquid clustering replaces table partitioning and zorder to simplify data layout decisions and optimize query performance it provides the flexibility to redefine clustering keys without rewriting existing data allowing data layout to evolve alongside analytic needs over time.

11:37 Feature engineering

So liquid clustering applies to both streaming tables and materialized views. So for us I think it's up to four columns of the partition keys cluster keys we can apply. So for us maybe sometimes we do it on date and sometimes we apply the keys on events. So right now you have a job scheduler you have your data easy to query and then it's the time for the real story right so I want to share a personal story but you guys can keep looking at the slide is last year I was doing a road trip in Maine if you guys don't know where that is close to Canada super cold December right so what is Maine famous for does any of them know that I hear lobster somewhere in Vermont.
Okay, that's a good answer though, but I don't agree. So, it was famous for lobster and I was craving one after road trip uh on a daily basis, right? So, but but I was driving an EV. I need to plug in for my day trip tomorrow. So, I went to Hannerford, which is like a local grocery store. I plug in there's a fast charging there, but the EV I was driving is not Rivian, just disclaimer. So, I I saw I saw the time was like 20 minutes.
So I thought, okay, I can go in there, get a lobster, and get it steamed and buy buy a bottle of Chardonnay and then just come back to my hotel and enjoy some Netflix. But when I came back, I paid everything like around 20 minutes. It still shows 20 minutes left. I was like, did I time travel or something? But no, right, it's because like it's not accurate and I have to base on that number to plan my trip. So that's why having a charge time estimation is important and my story kind of covers all the factor here if you see.
So charge time estimation is uh made of like consist of a holistic factors that we can see in the vehicle includes vehicle attributes including battery voltage, current, temperature and the state of charge. Uh also environment factors like temperature and finally the charger um factors like what is the pin temperature, what is the charge current limit, stuff like that. So the reason that in my case it was not that amazing is the system needs to ramp up at the beginning of the charging and you know a lot of the prediction it gives is not that accurate.
So we want to include all these holistic factors including the battery temperature you know like the environment scenarios and also the charger situation all together give the user a more accurate charge time estimation. So feature engineering is very important. We need to understand our data. You need to do EDA. You need to make sure that the columns we are importing into the model is relevant to this issue. And then right now we have done data processing.
Next up following the beautiful yellow RNS we are going to do model training. So every time we have a machine learning problem, we need to make sure that the data we input is going to be like the data we use or this project is viable, right? We don't want to just deep dive in it, work on something, use irrelevant data or some model and then there is like the evaluation metrics or the performance of model is not good.

15:02 Model training and tuning

So for this agile and fast test proof of idea we utilize AutoML. Basically we can import our data set in there and see define our target column and see like whether the output is good. So I'm just walking through like AutoML like to see how it works. So after define the experiment you're able to see this kind of report. So on here this clicker is working fine. Okay. So you can see you know how long you wait let's say two hours and then what's your evaluation metrics and then what's the run name what is the data set use for the model what model it uses and what's the evaluation metrics over there right most importantly we will have this view notebook for best model and also those notebook so we are able to utilize those notebook because the notebook has a lot of functions basically all the functions from what I listed there is in this my script and I find is super helpful. uh but a lot of those could be recycled if you have let's imagine because this will just be specific to just one topic right if I have more projects or I want to test another model using different data then I have to tweak it myself but to make this whole process more agile we have this function modularization process where we take a lot of functions in the script that could be recycled reused and then make into utilities uh library so it includes prop uh pre-processor utilities ities uh library patching utilities um feature importance and logging utilities.
So after we modularize all those functions they are very easy to query and easy to define. So we can make this whole thing very generalizable to other projects or other models and for other projects. Let's say you have new model you can just simply call that and redefine our own training testing data split our own pre-processor logic our own objective function. So now going into training because uh from the autom we have a general idea okay the project is working the evaluation metrics from the output is great and we know you know what's the initial standpoint of the hyperplanner space we can have there but if you want to make the model even better uh we need to do some DIY ourselves first most important factor is hyperparameter tuning so on the top right you will see there are like three basic strategies first one is great search basically just loop through all the combinations and then find find the minimum value the optimized value then second one is random search and third one is bashian uh uh bash optimization and for the advanced uh advanced uh strategy for hyperbrand tuning we have hyper optuna and retune so I have listed some of the pros and cons there maybe like you can have a broad understanding of if you are working in different industries which one is the best for you so when comparing hyperop opt Tuna and Retune each has its strengths depending on your use case.
Right. Hyper opt is one of the earlier two and supports search algorithms like TPE and random search. It has basic pruning support and is relatively simple to use make it great for quick straightforward experiments. Optuna builds on that with a more modern design. It's very userfriendly and supports advanced search methods like multi-objective optimization. It also comes with built-in pruning and visualization tools which makes it ideal for research and prototyping.
On the other hand, retune is built for scalability. It supports many search and pruning strategies including those from hyper opt and optuna and allows for seamless distributed parallelism across CPUs, GPUs and clusters. It is most flexible but requires a bit more setup. So in short, if you have a simple job, maybe you can go for hyper opt. If you want to just do it fast and for experimental optuna is your friend. And then finally if you want to scale your tuning into production level maybe you can try retune.
That's for hyperparameter tuning and we will have our you know hyperparameter uh space defined. So next step we'll go into the deep uh deep dive model training. So this is the whole pipeline layout. I'm just going to quickly go through it. The first one is column selector. That means you know maybe in the system you export a thousands of signals features to your model but your model needs to know which features are relevant to this issue right the model needs to know you go into the system which columns you are going to grab so that's the column selector and then you will have the transformers basically you have different data types numerical data categorical data boolean or date it needs to be transformed into certain format for model train finally combining all those transformers we have the pre-processor and then we can dump the data into machine learning model and then we do hyperparameter tuning and we try to optimize the objective function to find our best model.
Yeah. on the on the right but it's always important for the machine learning development is you have to save all the artifacts like what the model is consisted of like the attributes right you need to save your that's why we are saving you know what's the feature importance of all the features what is the uh training data what is the validation data and those will be reused in the future and we are going to talk about that in the following slides so now we are done with model training we can run our experiment and train our model.
So next step following the our we have model registration. So here is a bird eye view of model registration. You can see on the very left we have always those experiments and what experiments mean is basically the slides I shown before is you have all those uh model trainings and for each experiment you have the hyperparameter space defined and under the experiment is going to try out all the combinations of different hyperparameters to see which model using the same data is going to give you the most optimal result.

20:50 Model registration

So for each experiment you have bunch of model logged in there and each of the model uh in using MLflow we are able to see what it what are the evaluation metrics how the model performs and then based on the evaluation metrics we define let's say some people are interested in MSE mean square arrow some are more interested in the map mean absolute percent a error so depending on what kind of objective function you define you can find your best model for each experiment.
So each experiment will output at this stage the best model right and the best model is selected from all those trails and then depending on your use case or you you test it on some data you can do some registration let's say you see this best model you want to register it you want to keep it so that's where the versioning comes into the play so you want to version these models in for this project for this specific task for us is charge time estimation we have like first experiment ment this hyperparameter space like the best model we have it that's version one and for others maybe we maybe we tweaked our input uh sorry we tweaked our feature data and we switched our training test strategies or we have different hyperparameter space that'll be version two etc.
So you you can have all those versions of model registered in data bricks environment but those are all the winning models right like those are best model from the experiments how do we select like which one do we decide on to be customer facing so we need to make sure that when we compare the model performance it's apple toapple comparison so we always use this kind of a batched uh aggregated data which are super helpful for us because it's always the data that the model like any of those model haven't seen before.
So that's the new data the model haven't seen and then the new data will be applied to each of the model. In this case a evaluation matrix for each of the best model will be ranked. So depends on are you care more about like kind of evaluation matrix you care the most or you care more about generalization or like some accuracy the raw accuracy of the model you can rank them and then you can find your prop model. So the Spongebob is pretty happy is cheering but I think it's streaming too early because we are just halfway through.
The first stage is model deployment. Okay. So right now from the previous step we saw the best model like the champion model from all the experiments right and we say okay it's ready to deploy. We have this prop model data bricks make it easy. So we can use the serving endpoint feature and we can select the first one is uh your entity. So what entity means is what model registry you you locked all your model in like uh and then it's basically a folder and then you can select all the versions that you saved in the previous step.

23:53 Deployment strategies

So after you find maybe version seven is the best then you just put version seven over there and then define a compute type and then assign concurrency. So for concurrency assignment it will always require some domain expert size or some insights from the fleet data because that means you know every second how many users do you expect to call your model? I say if you run a social media company and you know you have thousands of users and or even millions of users each second you may have thousands of users using this service.
So you really need to scale up your concurrency but you you're not sure like is it is it going to be enough or or you're not sure like how many user is going to use it. You can start with a small number. So in this case we are able to deploy our you know champion of the champion model into this serving endpoint. And the model will be there and then we need to understand like for future monitoring you need to make sure that we are able to see what is the input what is the output of the model.
So in this case uh we always activate this function to enable the inference table and in this will give us the flexibility and also availability of the like what the user inputs into the model and what the model actually predicts. You just need to define your own schema of the data like at the very bottom on the right on the right you are able to see what kind of attributions you can define using this service uh you you will get from the ref the live reference table uh sorry the live inference table includes you know use uh influence like what date the model is called what's the time stamp of the model is called what are the request input features and what are the output features of the model so having all this data information uh like uh set up we are able to retrieve you know our model for uh performance like for future monitoring work and finally I mentioned sometimes you care more about accuracy sometimes you care more about generalization so in this case you you want to do AB testing you you're not sure which model to deploy so you're able to do the traffic split for AB testing in here you can add a serve entity and basically assign different traffic on those blocks saying some if you have two models maybe each one 50% and then the downstream inference table you'll be able to see what is the model that you are using when the user calls it and uh also like you are able to query some of the models if you want to do a baseline if you're working language model if you do want to do a baseline comparison you can always compare your own fine-tuned model with other baseline model there to see whether your fine tune model is actually improving Okay, so after you have this whole setup of serving endpoint and it's there and customer starts using or is in shadow mode, we are able to monitor what is the latency of this model. basically means uh how long the user has to wait before they like the event is triggered and then they get the response from the model.
And on the right you are able to see what's the request rate and based on the request rate you can go back to concurrency um definition to see do you need to increase it do you need to decrease it because like if you have your concurrency set up to a very high number it's extra cost but if you see okay it's not actually that commonly used maybe you can shrink it down a little bit to save the cost. So that's a cloth pathway like for ser endpoint deployment you have model deployed on the server endpoint then we need to make this communication between the vehicle and the endpoint right and this diagram shows um bird eye view of how we how we implement uh this communication so our definition is always when when the vehicle is plugged in and charging starts uh the vehicle will be able to know what kind of feature it will look into so we aggregate all those features uh within a certain time And if there's any null values because in the vehicle there could be situations happening and if there's any null values we have this uh fleet generated and then reference data which will be used to fill in the null value and then you have this complete data set you uh convert it into a JSON format and use the rest API called the ser endpoint API uh the ser endpoint model and the output input will be registered into datab bricks inference table and then also to uh to the vehicle and to the user.
So this is for cloud pathway deployment and some of you guys may ask okay because Rivian our commitment is to keep the world open forever right and what if like you're in the wood you're isolated you don't have internet what does user still have this service so for that's what we use to resolve it to age deployment so to deploy machine learning models efficiently on age devices you need a lightweight and interoperable across platform service which is onyx or open your network open your network open your network exchange.
So it serves exactly the purpose of the lightweight and like flexible service. It provides a standardized format to represent models trained on various frameworks like pytorch tensorflow tensorflow or xg boost. so they can be executed on resource constraint environments such as embedded systems, mobile services or in vehicle controllers. Converting a model to Onyx enables hardware acceleration, reduces inference latency and ensures portability across different runtimes.
This step is very important in production pipelines where low latency and platform independence are key to real-time decision- making at the edge. So how do we do it? Our next slide shows you know our model is not just a regressor, right? And uh as I mentioned above like there's a whole p there's a complete pipeline this consisted of the column selector uh the function transformer and like inside of it you have the imputer the standard scaler and then the regressor.
So for onyx conversion you have convert everything and that's circle back to the point where we say okay when you register a model you have a model you need to make sure you have everything that's that model is consisted of available that's a training data validation data feature importance table etc. So we are able to retrieve the training uh data that we used to train the cloud model and basically imitate this whole process for each of the um previous steps in the pipeline and then refit those steps into onyx format and plus the regressor onyx conversion we are able to convert this whole this whole pipeline into onyx uh format.
So in this case we have our cloud pathway defined we also have our age uh pathway defined then finally exciting stuff monitoring you have our model deployed you want to see how it works right so what are we monitoring most importantly we want to uh supervise all those drift data drift label drift and how those how those are caused because you know the seasonality when we maybe when we train a model is warm And when actually like in the winter it's cold the distribution of data being trained model is different and there could be you know a new emergence of new charging network which becomes like very fast be the dominant and then basically the charging profile is different so the input features distribution are different and maybe our user behavior is different like a lot of user starts to have more top of charge may they charge 100% which will be relevant to the data charge time performance.

31:30 Monitoring and retraining

So data drift, label drift is very important. Label drift is how our target variable is the predictor could change and that can be caused by on the right you'll see the upstream software update including you know there's a battery performance update uh or the current profile is changing or there's like a new vehicle platform or new battery pack. So combining all these two together we need to make sure our model is generating um stable and accurate statistics.
So fleet prediction uh offset statistics statistics is the one that we are actually monitoring to make sure okay what is the prediction offset across a week or across month across certain time frame and then you also monitor model latency you know how fast user get a response is there more lag than before and then there's maybe software and model input output discrepancy because maybe what model outputs there and going down to the software there's a misalignment there and we need to understand why and finally how often do we see that no connectivity issue.
So based from this monitoring system, we do have some real life challenges and this is a real case. So if you guys remember this plot is what I showed before on the uh medallion architecture, right? And this data is all used for training data and they're all based on cloud. But for this training model, it's not always perfect because it's different environment like cloud environment where data we receive could be different from the environment in the vehicle. to resolve it.
Uh like cases where vehicle signal have different sampling rate or not enough data and getting populated. We are get we are utilizing the inference table in data bricks because that's what actually in the field the input will be. So we concatenating what we see on the fleet into a new set of training data plus the cloud data because cloud data still have massive value. So combining those together we have this improved training data and then this improved training data will go down into this uh cyclic development through data processing model training registration deployment and then we need to monitor it again.
So with all that with all those in mind we have our exciting results. The first one definitely we are having this accuracy enhancement by at least 10%. And second of all uh which is we are super proud of is this machine learning project is driven how the product is designed right so we are able to give the flexibility of the user to understand how long it's going to take for them to charge to different steps of uh their um of their charge bar.

34:28 Results and future outlook

So before it was only you know you have setting to 80% then you only give you one number but right now we can understand if your setting is 80 but if you ch but but we can give you the output is saying if you charge with 60 to 50 like how long it's going to take so you can better plan your route and then we have we provide uh a significant amount of internal insights on higher granularity charging behavior and on features beyond physics based simulation model and also if there's anything happens in the model we see on our back end there's any error.
The integration is very agile. It's very fast. So the simulus integration we can do it with minimal disrupt disruption. And here's the actual product. This is Rivian energy app 2. 0. On the left is what you see on the charging panel. Okay. So it's idle. So you're not charging actually. You can see here what is your current SOC state of charge and what is your setting if you start to charge. But on the right, if you plug in to a charging station, you're able to see where you stand.
That's a green bar. And also the gray bar is what uh the future charging will happen and how long it's going to take. Also, this curve can tells you what is the energy output input from the vehicle and the charger. So, this is really exciting and uh the result is good. I'm going to hand it over to Mo to wrap it up. Awesome. Thanks, And so what is next? We want to scale our edge deployment. Um I think this is a um a great example of a problem where physics based and AI and machine learning together could develop a hybrid approach that could solve one of the problems that's related to batteries charging and overall EV adoption.
So edge deployment is a great opportunity for us to tackle. Second is using that workflow that has been already set up and scaling it to other problems as well is our second um step. Third is that um we're growing the team within the company. So for all of you guys, please feel free to scan the QR code and join our team. Okay, that's a disclaimer that the content presented here is uh forformational purposes only and for more accurate um and current information, please refer to the official Rivian communications.
Uh so yeah, finally, please take uh a moment to fill in your survey. Uh thank you. Any questions?

Learn more about the Databricks Data and AI platform.

The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.