It is rarely necessary to transform data. Parametric statistical models don't have assumptions about the distribution of data (only the distribution of model residuals; it's possible to have a non-normally distributed independent or dependent variable but a regression model that still yields normally-distributed residuals).

If the residuals fail to be normally distributed, this is often because there is some other important covariate you have not accounted for (see, e.g., examples in https://rpubs.com/iabrady/residual-analysis), so the solution is usually to include the missing covariate, rather than transforming the data.

If you do eventually conclude that transforming the data is the only way, there are many possible transforms available (square root, natural log, log10, reflected reciprocal), and you should keep in mind that the model and coefficients may become harder to interpret since they'll be expressed in transformed units rather than original units. You'll want to check which transform yields a model with the most normally-distributed residuals. I usually prefer to use a transform where it's easy to recover the original units (e.g., log) so the unstandardized coefficients and confidence intervals can be transformed back into original units.

Hello dr Shakil ,,,

It is very good question. Added to the previous answer, this technique is being used in six sigma methodology to improve process performance.

https://www.isixsigma.com/tools-templates/normality/making-data-normal-using-box-cox-power-transformation/

Six sigma methodology done by minitab software

https://blog.minitab.com/blog/applying-statistics-in-quality-projects/how-could-you-benefit-from-a-box-cox-transformation

Yes, I feel both the answers are really informative and explained in lucid way.