Non-dimensionalization of variables to enhance machine learning in additive manufacturing processes
A method for training a machine learning engine for modeling of a physical system includes receiving process data representing measurements of a physical system. The method includes applying a transform to values of the at least two variables of the process data to generate a dimensionless parameter having a parameter value corresponding to each measurement of the physical system for the at least two variables. The method includes training the machine learning engine using a set of generated training data including the non-dimensionalized parameter, to output a prediction of a value of a physical effect of the physical system for values of the variables that are not included in the process data. The method includes controlling an additive manufacturing process for the material by setting the at least one physical property to the value of the at least one process variable during fabrication of a part.
1. A method for controlling an additive manufacturing process, the method comprising:
receiving a specification of at least one physical property of the additive manufacturing process for a material;
determining, from the specification of the at least one physical property, a process outcome for fabrication of a part from the material;
retrieving a process map comprising a set of process variables that are configured to cause the process outcome, the process map being generated by:
receiving a set of training data comprising a dimensionless parameter, the dimensionless parameter being a combination of at least two variables each representing a physical property of the additive manufacturing process; and
applying the set of training data to machine learning logic to generate predictions of the process outcome for different values of the dimensionless parameter, the predictions of the process outcome from the process map, wherein the machine learning logic comprises neural network regression logic;
determining, from the process map, a value of at least one process variable for the additive manufacturing process to cause the process outcome in the material; and
controlling the additive manufacturing process for the material by setting the at least one physical property to the value of the at least one process variable during fabrication of the part, wherein controlling comprises:
determining, based on the process map, a particular value for the at least one process variable for fabricating the part from the material; and
causing fabrication of the part form the material based on the particular value for the at least one process variable.
2. The method of claim 1 , wherein generating the process outcome is based on one or more of a melt pool geometry, a temperature gradient in the material, a temperature integral in the material.
3. The method of claim 1 , wherein the generated process outcome comprises one of a dimension of the part or a surface roughness of the part.
4. The method of claim 1 , wherein controlling the additive manufacturing process further comprises:
determining, based on the process map, the particular value for the at least one process variable for causing a melt pool to evaporate at a specific rate;
adjusting heat in a region of the part based on the particular value; and
causing the melt pool in the part to evaporate at the specific rate.
5. The method of claim 1 , further comprising updating the process map by:
receiving values of measurements of an additive manufacturing process, each measurement corresponding to a variable for the additive manufacturing process;
receiving data indicating a relationship between at least two of the variables;
applying a transform to the values of the measurements of the at least two variables to generate additional values of the dimensionless parameter;
updating the set of training data for the machine learning logic based on the additional values of the dimensionless parameter; and
training the machine learning logic using the updated set of training data.
6. The method of claim 1 , wherein the dimensionless parameter is generated by:
selecting, based on the specification, at least two process variables of the additive manufacturing process that are physically related; and
combining the at least two process variables into the dimensionless parameter.
7. The method of claim 1 , wherein the dimensionless parameter is generated by:
selecting, based on the specification, at least one process variable of the additive manufacturing process and at least one material variable of the additive manufacturing process, the at least one process variable and the at least one material variable being physically related; and
combining the at least one process variable and the at least one material variable into the dimensionless parameter.
8. The method of claim 1 , wherein the dimensionless parameter comprises at least one of:
η
=
qv
4
π
α
k
(
T
m
-
T
0
)
,
a
=
Av
2
4
α
2
,
w
=
Wv
2
α
,
and
η
=
qv
4
π
α
k
(
T
m
-
T
0
)
,
wherein q is a beam power, v is a beam velocity, α is a thermal diffusivity of the material, k is a thermal conductivity of the material, T m is a melting temperature of the material, T 0 is an initial temperature of the material, A is a melt pool area, a is a dimensionless parameter corresponding to the melt pool area, W is a melt pool width, w is a dimensionless parameter corresponding to the melt pool width, D is a melt pool depth, and d is a dimensionless parameter corresponding to the melt pool depth.
9. The method of claim 1 , wherein the set of training data comprises measurements only for materials other than the material of the additive manufacturing process.
10. The method of claim 1 , wherein the at least one physical property comprises one of a type of the material, a feed type of the material, and a type of a heat source for heating the material.
11. The method of claim 1 , wherein the at least one process variable is selected from a group comprising a power (P) variable associated with a thermal process, a translation speed (V) variable associated with the thermal process, a material feed rate (MFR) variable (or variable related to MFR) used in the thermal process, one or more structure geometry variables, and a structure temperature (T 0 ) variable.
12. The method of claim 1 , wherein generating the process map comprises interpolating temperature integrals based on applying the set of training data to the machine learning logic.
13. The method of claim 1 , further comprising generating the training data by:
performing a series of tests at different values of the at least one physical property; and
measuring a melt pool geometry in the material for each of the different values of the physical property.
14. The method of claim 1 , wherein the process map is generated from the set of training data including values for a first alloy, and wherein the process map includes one or more predictions of a melt pool geometry for causing the process outcome in a second alloy that is different from the first alloy.