Page 53 - FDMAsia NovDec 2024

MATERIALS 51
www.fdmasia.com | FDM ASIA NOV/DEC 2024

Results network in order to see if it is possible to predict useful
For the characterisation of the model, it was preferred to information such as a deflection for a beam in the case of
generate all possible dimensional combinations of beams (Table the topologically optimised elements, without having to run
1), respectively, to vary the dimensions as presented below: the topological optimisation. This represents a tried and tested

technique for civil engineering applications. The architecture
Table 1. Dimensional combination of beams. of the deep neural network is described.
Type of Element Dimensions Pitch A total of 6526 topologically optimised models/datasets
Beam length (L) from 4 to 8 m 10 cm (pairs of synthetic parameters—displacements) was divided
Beam width (W) From 65 cm to 1.65 m 5 cm into two parts: 6388 models for ANN training and 138 models
Beam span from 3 to 5 m 50 cm for validating the result. The ANN optimisation algorithm was
Thickness of a plank 4 cm the SGD (Stochastic Gradient Descent) algorithm.
Length of the support area 20 cm, 30 cm and 40 cm The ANN model thus obtained shows a good alignment
both from the point of view of predictions for training data and

Following the combinations in the model resulted in 6526 from the point of view of predictions for the set dedicated to
geometrically distinct patterns. Running these models required validation where the prediction of the ANN model is represented
over 24,000 h/core, which was executed by outsourcing the on the horizontal axis, and the actual value is displayed on the
service. vertical axis. Some edge-cases were also showcased where
Each of these patterns were statically analysed in the the model tends to underpredict the actual deflection; this is
Abaqus 2023 program with the following optimisation why there are more points above the 100 percent alignment
strategies: line, which can be improved by further developing the ANN

• Planar symmetry with respect to the vertical plane architecture.
perpendicular to the longitudinal axis of the beam in the After finalising the training, a benchmark was run over
centre of gravity; the training dataset (6388 models) in order to benchmark
• Planar symmetry with respect to the horizontal median how much faster the neural network is in predicting the
plane; deflection than running the FEM optimisation task using the
• Polar symmetry with respect to the centre of gravity of timeit Python (3.13.0) library. The results obtained indicate that
the beam. this information can be obtained in 1.4 milliseconds using the
ANN (best result out of 10,000 runs).
Each optimiser was run with 100 iterations and the final The evolution of the objective function value over the training

iteration was processed. epochs was shown. In this case, the objective function value
Due to the large space occupied by the results of a was considered to be the sum of the Mean Squared Loss
turnover, only the useful synthetic information presented in value for each data point, over one training epoch.
the Annexes was saved.
In order to test the feasibility of embedding the topological Optimised Geometric Solutions
optimisation information into a machine learning model, the The current project aims to start the development of an expert
first step is to create a regressive model and to test how it software tool, based on Artificial Intelligence, which will provide

responds to the input parameters and objectives. optimised geometric solutions for glulam elements.
The approach chosen was to develop a deep neural The formation of the database for training the Artificial