Making Predictions with GlassPy

Author: Daniel R. Cassar

Introduction

GlassPy provides three prediction models: GlassNet, VITRIFY, and ViscNet.

GlassNet is a multitask deep neural network that predicts 85 glass properties and the temperature dependence of viscosity.
VITRIFY is a classification model that estimates the probability of a given composition forming a glass.
ViscNet is a physics-informed deep neural network designed specifically to predict the temperature dependence of viscosity.

All three models were trained on data from the SciGlass database.

GlassNet basic usage

Below is a minimal example of how to load and use the GlassNet model.

[1]:

from glasspy.predict import GlassNet

model = GlassNet()
composition = "Li2O(SiO2)2"
predictions = model.predict(composition)

predictions

[1]:

	T0	T1	T2	T3	T4	T5	T6	T7	T8	T9	...	Cp1673K	TMaxGrowthVelocity	MaxGrowthVelocity	CrystallizationPeak	CrystallizationOnset	SurfaceTensionAboveTg	SurfaceTension1173K	SurfaceTension1473K	SurfaceTension1573K	SurfaceTension1673K
0	1612.586818	1286.966695	1187.265614	1070.224145	968.06795	901.811624	859.579252	825.204682	782.490883	759.993951	...	1717.653045	1040.284178	-6.241866	937.262437	788.40385	0.316586	0.3044	0.321539	0.318579	0.319566

1 rows × 85 columns

A composition can also be defined using a dictionary.

[2]:

composition = {
    "SiO2": 2,
    "Li2O": 1,
}
predictions = model.predict(composition)

predictions

[2]:

	T0	T1	T2	T3	T4	T5	T6	T7	T8	T9	...	Cp1673K	TMaxGrowthVelocity	MaxGrowthVelocity	CrystallizationPeak	CrystallizationOnset	SurfaceTensionAboveTg	SurfaceTension1173K	SurfaceTension1473K	SurfaceTension1573K	SurfaceTension1673K
0	1612.586818	1286.966695	1187.265614	1070.224145	968.06795	901.811624	859.579252	825.204682	782.490883	759.993951	...	1717.653045	1040.284178	-6.241866	937.262437	788.40385	0.316586	0.3044	0.321539	0.318579	0.319566

1 rows × 85 columns

GlassNet also accepts pandas DataFrames as input. Note that each row represents a material and that only columns related to chemical composition can be present in the DataFrame.

[3]:

import pandas as pd

data = [
    [1, 0, 2],
    [0, 1, 2],
    [1, 1, 2],
]

df = pd.DataFrame(data, columns=["Li2O", "Na2O", "SiO2"])
df

[3]:

	Li2O	Na2O	SiO2
0	1	0	2
1	0	1	2
2	1	1	2

[4]:

predictions = model.predict(df)

predictions

[4]:

	T0	T1	T2	T3	T4	T5	T6	T7	T8	T9	...	Cp1673K	TMaxGrowthVelocity	MaxGrowthVelocity	CrystallizationPeak	CrystallizationOnset	SurfaceTensionAboveTg	SurfaceTension1173K	SurfaceTension1473K	SurfaceTension1573K	SurfaceTension1673K
0	1612.586818	1286.966759	1187.265614	1070.224201	968.067993	901.811569	859.579252	825.204682	782.490883	759.993951	...	1717.653120	1040.284178	-6.241865	937.262560	788.403850	0.316586	0.304400	0.321539	0.318579	0.319566
1	1700.171170	1521.813466	1287.576204	1130.354623	1036.780176	955.844950	911.027803	848.791035	810.406750	782.243953	...	1482.022135	1058.443194	-6.394961	963.434464	905.734703	0.279031	0.310337	0.290839	0.287974	0.272271
2	1457.627094	1246.016165	1114.455164	987.722016	934.740619	861.280868	794.972760	769.729834	730.195119	700.667308	...	1751.036502	1082.912718	-5.101020	875.732689	820.388535	0.299141	0.308964	0.319907	0.321466	0.294990

3 rows × 85 columns

GlassNet can also predict viscosity and the MYEGA viscosity equation parameters.

[5]:

predictions = model.predict_log10_viscosity(
    T=1000,
    composition=df,
)

predictions

[5]:

array([3.53501871, 4.40572086, 2.83984945])

[6]:

viscosity_parameters = model.viscosity_parameters(df)

viscosity_parameters

[6]:

	log10_eta_infinity (Pa.s)	Tg_MYEGA (K)	fragility
0	-1.222603	701.790333	42.844319
1	-1.660096	710.677354	35.863589
2	-1.494931	647.762409	40.367431

VITRIFY base usage

Below is a minimal example of how to load and use the VITRIFY model. Like GlassNet, VITRIFY accepts strings, dictionaries, and pandas DataFrames as input. It returns 1 if the composition is predicted to form a glass under typical laboratory conditions, or 0 otherwise.

[7]:

from glasspy.predict import VITRIFY

model = VITRIFY()
predictions = model.predict(df)

predictions

[7]:

array([1, 1, 1], dtype=int64)

You can also retrieve the probability that the composition will form a glass, as estimated by the model.

[8]:

probabilities = model.predict_proba_glass(df)

probabilities

[8]:

array([0.96919906, 0.99467885, 0.73446614])

ViscNet basic usage

ViscNet is used similarly to GlassNet. However, since GlassNet outperforms ViscNet on viscosity prediction, using GlassNet is generally recommended for that task. The minimal example below demonstrates how to load and use ViscNet.

[9]:

from glasspy.predict import ViscNet

model = ViscNet()
log10_viscosity = model.predict(T=1000, composition=df)

log10_viscosity

[9]:

array([5.197888 , 5.6581345, 6.0721517], dtype=float32)

[10]:

fragility = model.predict_fragility(df)

print(fragility)

[34.436737 29.760372 34.16933 ]

[11]:

Tg = model.predict_Tg(df)

print(Tg)

[748.3823  739.4691  788.51385]