Prediction of mechanical and penetrability properties of cement-stabilized clay exposed to sulfate attack by use of soft computing methods


SEZER A., İNAN SEZER G., MARDANI AGHABAGLOU A. , ALTUN S.

NEURAL COMPUTING & APPLICATIONS, vol.32, no.21, pp.16707-16722, 2020 (Journal Indexed in SCI) identifier identifier

  • Publication Type: Article / Article
  • Volume: 32 Issue: 21
  • Publication Date: 2020
  • Doi Number: 10.1007/s00521-020-04972-x
  • Title of Journal : NEURAL COMPUTING & APPLICATIONS
  • Page Numbers: pp.16707-16722
  • Keywords: Cement-stabilized soil, Strength, Penetrability, BPNN, ANFIS, Soft computing, STRENGTH DEVELOPMENT, BEHAVIOR, DENSITY

Abstract

Similar to its effects on any type of cementitious composite, it is a well-known fact that sulfate attack has also a negative influence on engineering behavior of cement-stabilized soils. However, the level of degradation in engineering properties of the cement-stabilized soils still needs more scientific attention. In the light of this, a database including a total of 260 unconfined compression and chloride ion penetration tests on cement-stabilized kaolin specimens exposed to sulfate attack was constituted. The data include information about cement type (sulfate resistant-SR; normal portland (N) and pozzolanic-P), and its content (0, 5, 10 and 15%), sulfate type (sodium or magnesium sulfate) as well as its concentration (0.3, 0.5, 1%) and curing period (1, 7, 28 and 90 days). Using this database, linear and nonlinear regression analysis (RA), backpropagation neural networks and adaptive neuro-fuzzy inference techniques were employed to question whether these methods are capable of predicting unconfined compressive strength and chloride ion penetration of cement-stabilized clay exposed to sulfate attack. The results revealed that these methods have a great potential in modeling the strength and penetrability properties of cement-stabilized clays exposed to sulfate attack. While the performance of regression method is at an acceptable level, results show that adaptive neuro-fuzzy inference systems and backpropagation neural networks are superior in modeling.