Enabling fast charging

What can we do?

Outline

Problem formulation
Modeling
Experiments
Results/Takeaways

Problem formulation

Fast charging

Problem statement

How could we enable fast charging that produces higher cycle life?

Problem statement

How could we enable fast charging that produces higher cycle life?
Scientists landed on few things we can control: biggest one is the electrolyte

Problem statement

How could we enable fast charging that produces higher cycle life?
Scientists landed on few things we can control: biggest one is the electrolyte
Conductivity of the electrolyte is a key factor

Problem statement

How could we enable fast charging that produces higher cycle life?
Scientists landed on few things we can control: biggest one is the electrolyte
Conductivity of the electrolyte is a key factor
How could we optimize/maximize the conductivity?

Problem statement

How could we enable fast charging that produces higher cycle life?
Scientists landed on few things we can control: biggest one is the electrolyte
Conductivity of the electrolyte is a key factor
How could we optimize/maximize the conductivity?
Do we have data?

Problem statement

How could we enable fast charging that produces higher cycle life?
Scientists landed on few things we can control: biggest one is the electrolyte
Conductivity of the electrolyte is a key factor
How could we optimize/maximize the conductivity?
Do we have data? Yes*

Data

Data [1235 x 131]

What does electrolyte contain?

Solvents

What does electrolyte contain?

Solvents
Solvents mol %

What does electrolyte contain?

Solvents
Solvents mol %
Salts

What does electrolyte contain?

Solvents
Solvents mol %
Salts
Salts mol %

What does electrolyte contain?

Solvents
Solvents mol %
Salts
Salts mol %
Additives

What does electrolyte contain?

Solvents
Solvents mol %
Salts
Salts mol %
Additives
Additives mol %

What does electrolyte contain?

Problem formulation - attempt

Create a regression model that uses all electrolyte ingredients along with their percentages to predict Conductivity (mS/cm)

Problem formulation - attempt

Create a regression model that uses all electrolyte ingredients along with their percentages to predict Conductivity (mS/cm)
Nudge experiments

Problem formulation - attempt

Create a regression model that uses all electrolyte ingredients along with their percentages to predict Conductivity (mS/cm)
Nudge experiments
Catch! we have interesting constraints.

Constraints

Famous Solvent % \(\leq\) 2* Famous Salt %

Constraints

Famous Solvent % \(\leq\) 2* Famous Salt %
Another Solvent % \(\leq\) 2* Famous Salt %

Constraints

Famous Solvent % \(\leq\) 2* Famous Salt %
Another Solvent % \(\leq\) 2* Famous Salt %
additive % \(\leq\) Fixed Percentage

Constraints

Famous Solvent % \(\leq\) 2* Famous Salt %
Another Solvent % \(\leq\) 2* Famous Salt %
additive % \(\leq\) Fixed Percentage
salt % \(\leq\) Another Fixed Percentage

Constraints

Famous Solvent % \(\leq\) 2* Famous Salt %
Another Solvent % \(\leq\) 2* Famous Salt %
additive % \(\leq\) Fixed Percentage
salt % \(\leq\) Another Fixed Percentage
sum % = 100

Constraints

Famous Solvent % \(\leq\) 2* Famous Salt %
Another Solvent % \(\leq\) 2* Famous Salt %
additive % \(\leq\) Fixed Percentage
salt % \(\leq\) Another Fixed Percentage
sum % = 100
cycle life \(\geq\) Fixed Cycle Life

Problem statement

Create a Smooth/Polynomial regression model that uses all electrolyte ingredients along with their percentages to predict/infer Conductivity (mS/cm)

Problem statement

Create a Smooth/Polynomial regression model that uses all electrolyte ingredients along with their percentages to predict/infer Conductivity (mS/cm)
Use constraints to optimize the fitted polynomial model

Problem statement

Create a Smooth/Polynomial regression model that uses all electrolyte ingredients along with their percentages to predict/infer Conductivity (mS/cm)
Use constraints to optimize the fitted polynomial model
Find the arguments that maximize the Conductivity

Problem statement

Create a Smooth/Polynomial regression model that uses all electrolyte ingredients along with their percentages to predict/infer Conductivity (mS/cm)
Use constraints to optimize the fitted polynomial model
Find the arguments that maximize the Conductivity
Suggest a grid that chemically makes sense as experiments

Process workflow

Assumptions made

\[ \arg\max_{\theta} f(\theta) \approx \arg\max_{\theta} \hat{f}(\theta) \]

Assumptions made

\[ \arg\max_{\theta} f(\theta) \approx \arg\max_{\theta} \hat{f}(\theta) \]

Under ‘nice’ conditions

Uniform Convergence
Unique Global Maximum

Modeling

Features

Assume we have

n solvents
m salts
k additives

That is 2(n+m+k) features.

Categorical columns

Solvents

NA, None, ““, N/A, NONE
10 common values: simplified levels (regex)
Combined others

Categorical columns

Solvents

NA, None, ““, N/A, NONE
10 common values: simplified levels (regex)
Combined others

Salts

NA, None, ““, N/A, NONE
4 common values: simplified levels (regex)
combined others

Numerical columns

30%, 30 %, 30.0, 0.3, 30 perc, 30 percent

Numerical columns

30%, 30 %, 30.0, 0.3, 30 perc, 30 percent
70% –> 70
25.6% –> 25.6
abc –> 0
NA –> 0
0.3 –> 30

Numerical columns

sum did not added up to 100

Numerical columns

sum did not added up to 100
scientists cleaned it up using their records

Raw Data

Constraints

External constraints

Famous Solvent % \(\leq\) 2* Famous Salt %
Another Solvent % \(\leq\) 2* Famous Salt %
additive % \(\leq\) Fixed Percentage
salt % \(\leq\) Another Fixed Percentage
sum % = 100

Constraints

External constraints

Famous Solvent % \(\leq\) 2* Famous Salt %
Another Solvent % \(\leq\) 2* Famous Salt %
additive % \(\leq\) Fixed Percentage
salt % \(\leq\) Another Fixed Percentage
sum % = 100

Internal constraint

cycle life \(\geq\) Fixed Cycle Life

Internal constraint: cycle life

After cleanup

273 row

Conductivity

Further Feature engineering

poly_rec <- recipe(conductivity ~., elyte_data) |> 
  update_role(id, new_role = 'id') |> 
  # bin additives
  step_other(additive, other = 'others_combined', threshold = 0.05) |> 
  step_zv(all_predictors()) |> 
  # create squared columns
  step_poly(contains('p'), degree = 2, options = list(raw = TRUE)) |> 
  # create interactions between all numerical columns
  step_interact(terms = ~ ends_with("poly_1"):ends_with("poly_1"))

Further Feature engineering

poly_rec <- recipe(conductivity ~., elyte_data) |> 
  update_role(id, new_role = 'id') |> 
  # bin additives
  step_other(additive, other = 'others_combined', threshold = 0.05) |> 
  step_zv(all_predictors()) |> 
  # create squared columns
  step_poly(contains('p'), degree = 2, options = list(raw = TRUE)) |> 
  # create interactions between all numerical columns
  step_interact(terms = ~ ends_with("poly_1"):ends_with("poly_1"))

Further Feature engineering

poly_rec <- recipe(conductivity ~., elyte_data) |> 
  update_role(id, new_role = 'id') |> 
  # bin additives
  step_other(additive, other = 'others_combined', threshold = 0.05) |> 
  step_zv(all_predictors()) |> 
  # create squared columns
  step_poly(contains('p'), degree = 2, options = list(raw = TRUE)) |> 
  # create interactions between all numerical columns
  step_interact(terms = ~ ends_with("poly_1"):ends_with("poly_1"))

Further Feature engineering

poly_rec <- recipe(conductivity ~., elyte_data) |> 
  update_role(id, new_role = 'id') |> 
  # bin additives
  step_other(additive, other = 'others_combined', threshold = 0.05) |> 
  step_zv(all_predictors()) |> 
  # create squared columns
  step_poly(contains('p'), degree = 2, options = list(raw = TRUE)) |> 
  # create interactions between all numerical columns
  step_interact(terms = ~ ends_with("poly_1"):ends_with("poly_1"))

Further Feature engineering

poly_rec <- recipe(conductivity ~., elyte_data) |> 
  update_role(id, new_role = 'id') |> 
  # bin additives
  step_other(additive, other = 'others_combined', threshold = 0.05) |> 
  step_zv(all_predictors()) |> 
  # create squared columns
  step_poly(contains('p'), degree = 2, options = list(raw = TRUE)) |> 
  # create interactions between all numerical columns
  step_interact(terms = ~ ends_with("poly_1"):ends_with("poly_1"))

Further Feature engineering

poly_rec <- recipe(conductivity ~., elyte_data) |> 
  update_role(id, new_role = 'id') |> 
  # bin additives
  step_other(additive, other = 'others_combined', threshold = 0.05) |> 
  step_zv(all_predictors()) |> 
  # create squared columns
  step_poly(contains('p'), degree = 2, options = list(raw = TRUE)) |> 
  # create interactions between all numerical columns
  step_interact(terms = ~ ends_with("poly_1"):ends_with("poly_1"))

Coefficients

Optimization [with extrenal constraints]

library(nloptr)

# inequality constraints
eval_g_ineq <- function(x){
  return (x[1] - 2*x[7])
}

eval_g_ineq2 <- function(x){
  return (x[5] - 6)
}

eval_g_ineq3 <- function(x){
  return (x[6] - 6)
}

eval_g_ineq4 <- function(x){
  return (x[8] - 2)
}

eval_g_ineq5 <- function(x){
  return (x[9] - 2)
}

# equality constraint (sum of percentages must be 100)
eval_g_eq <- function(x){
  return ( sum(x)-100 )
}

Experiment suggestions

Experiments

Measured Conductivity

Was the experiment successful?

Results/Takeaways

Collect better CL data

Results/Takeaways

Collect better CL data
Optimize CL along with Conductivity

Results/Takeaways

Collect better CL data
Optimize CL along with Conductivity
ML model just as a predictive tool is valuable on itself

Results/Takeaways

Collect better CL data
Optimize CL along with Conductivity
ML model just as a predictive tool is valuable on itself
New business idea: Could we create ML platform?

Initiation of ML platform

Thanks!