Next-generation non-linear and collapse prediction models for short to long period systems via machine learning methods
The machine learning approach: Exterme Gradient Boosting (XGBoost)
Makes predictions for a strength ratio - ductility - period relationships
Key arguments:
-
$R$ - strength ratio based on spectral acceleration -
$\rho$ - strength ratio based on average spectral acceleration -
$\mu$ - ductility -
$T$ - period
where
-
$Sa(T)$ stands for spectral acceleration at fundamental period -
$Sa_y$ stands for spectral acceleration at yield -
$Sa_{avg,2}(T)$ stands for average spectral acceleration computed at periods$∈ [0.2T:2T]$ -
$Sa_{avg,3}(T)$ stands for average spectral acceleration computed at periods$∈ [0.2T:3T]$
pip install xgb-rhomut
Example 1: Dynamic strength ratio prediction of non-collapse scenarios at a dynamic ductility level of 3.0:
import xgbrhomut
model = xgbrhomut.XGBPredict(im_type="sa_avg", collapse=False)
prediction = model.make_prediction(
period=1,
damping=0.05,
hardening_ratio=0.02,
ductility=4,
dynamic_ductility=3.0
)
Example 2: Dynamic ductility prediction given a strength ratio of 3.0 (since im_type is "sa_avg", and collapse is False,
import xgbrhomut
model = xgbrhomut.XGBPredict(im_type="sa_avg", collapse=False)
prediction = model.make_prediction(
period=1,
damping=0.05,
hardening_ratio=0.02,
ductility=4,
strength_ratio=3.0
)
prediction:
{
"median": float,
"dispersion": float
}
Other methods
xgbrhomut.r_mu_t.ec8.strength_ratio(mu=3, T=1, Tc=0.5)
Limitations in terms of input parameters are:
-
$T$ ∈ [0.01, 3.0] seconds -
$\mu$ ∈ [2.0, 8.0] -
$\xi$ ∈ [2.0, 20.0] % -
$a_h$ ∈ [2.0, 7.0] % -
$a_c$ ∈ [-30.0, -100.0] % -
$R$ ∈ [0.5, 10.0]
where
-
$T$ stands for period -
$\mu$ stands for ductility -
$\xi$ stands for damping -
$a_h$ stands for hardening ratio -
$a_c$ stands for softening ratio (necessary to compute fracturing ductility, where collapse is assumed)
Predictions made using the non-linear analysis resutls of 7292 unique SDOF systems amounting in total to 26,000,000 observations (collapse + non-collapse).
- Shahnazaryan D., O'Reilly J.G., 2023, Next-generation non-linear and collapse prediction models for short to long period systems via machine learning methods, Engineering Structures 306(117801): 117801. https://doi.org/10.1016/j.engstruct.2024.117801